Untrusted Input

Memory-augmented and tool-using agents expose exact private values when remote LLMs process retrieved memory, tool actions, and intermediate observations. One-way masking limits direct exposure but removes values needed for trusted execution and can leak them through later observations. We propose Trustworthy Virtual Memory (TVM), a closed-loop runtime that keeps exact-value state local while presenting a protected view to the remote model. Within this single runtime, Rule-TVM replaces whole protected fields with locally recoverable handles, and Semantic-TVM instead replaces only sensitive spans predicted by a trusted local model, preserving surrounding task-relevant context. On Memory-EHR and Memory-RAP across two providers, span-level projection recovers most of the EHR utility lost under whole-field replacement (Task Success 84.17% vs. 52.33% on DeepSeek) while measured exposure stays low and workflows remain executable.

Multi-tenant tools commonly accept a tenant identifier and validate it against the caller's entitlement. For a large language model (LLM) agent, that pattern delegates resource selection to a process whose context may contain attacker controlled instructions. We formalize this stochastic deputy problem and present a structural defense: remove tenant identity from the Model Context Protocol (MCP) tool schema, bind scope to a verified credential, and enforce it below the agent. In a 373-trial ablation across eight model configurations and two transports, a correctly validated tenant parameter served every out-of-scope attempt: 26 of 26, or 26 of 41 plausible-pretext trials overall. With the parameter removed, no tool signature could express the read. Twelve of 56 trials instead escaped the interface by forging writable scope, showing that interface invariance requires cryptographically protected context. On a production dataset containing multiple GBs of data, set-valued scope caused a measured $57\times$ latency ratio under function-wrapped membership predicates; a JSON_TABLE lateral join recovered index access where the tenant key was indexed. The evaluation also exposes deployment limits, including an entitlement-size query-planner cliff and incomplete index coverage. The result is a tenant-isolation argument that depends on enforceable interfaces and credentials rather than model compliance.

Tool-augmented language models are evaluated on whether they reach the right answer, not on whether they report honestly when a tool fails to supply one. We isolate this post-failure decision with a benchmark of 1,024 items spanning 16 internal-system domains and eight tool-failure types, in which a tool call is enforced and the returned payload is guaranteed to be unusable. Under a deployment-style system prompt, 14.10% of responses are dishonest: the model either asserts a value the payload cannot support or declines while citing a fabricated policy or capability limit. The rate is governed almost entirely by whether the failure is signalled. When the tool returns status:error, dishonesty is absent (0.0%); when it returns status:ok with a redacted, corrupted, stale, malformed, empty or truncated value, dishonesty reaches 45.3%. The behaviour is not an artefact of our prompts: it appears under a neutral prompt (10.17%) and under the shipped prompt of every production agent framework we evaluate, reaching 24.67% under CrewAI's, and none of the nine frameworks we audit specifies what the model should do when a tool fails. Comparing prompt-level defences, we find that the operative variable is not deference to tool output but the absence of a named failure state. Appending a single sentence that requires the model to emit retrieval_status: OK or FAILED before answering reduces dishonesty from 14.10% to 0.87%, with one item of 688 worsening against 92 improving, and transfers unchanged into three foreign agent scaffolds. The emitted flag is faithful in 99.7-99.9% of declarations, giving a runtime detector that needs only a regular expression.

Tool-using language agents can delegate and revoke permissions while acting through external services. We show that two authorization histories can have identical current permissions and identical all-pairs reachability yet require opposite decisions after the same direct-edge revocation. We formalize the information needed to preserve such distinctions as a residual authorization state. We prove that exponentially many future-distinct states can share one fixed transitive closure, and give exact or tight asymptotic bounds on the state required by an exact monitor as delegation redundancy varies. ResidualAuth compiles these constructions into paired language-agent episodes. Across four open-weight models, a fixed 256-token summary solved 0-2/16 pairs, sham reads solved 0/16, and authenticated current-query reads solved 15-16/16. In a separate held-out online-memory diagnostic, exact ledger serializations fit all 128 four-coordinate pairs at both 768 and 1,024 tokens. At either cap, factually supported model-written memories sufficient for every prespecified continuation solved at most 1/128 pairs per model. A hard gate reduced eight observed unauthorized effects to zero without changing the preceding attempts. These results distinguish required authorization state, usable decision information, online state maintenance, and effect mediation.

Long-running LLM agents rely on persistent memory to carry state across interactions, including permissions, restrictions, and revocations. When memory misrepresents this evolving authorization state, the agent's own records can grant authority that the underlying history never permitted, resulting in misaligned behavior without any external attacks. We term this failure endogenous authorization laundering, where spurious permissions written into memory lead to unauthorized actions as their provenance is washed away. We then introduce EAL-Bench, which measures how accurately persistent memory preserves evolving authorization state and whether errors propagate to downstream unauthorized actions. We evaluate five LLMs as memory writers and two as executors across procurement, cybersecurity, and finance. We find that under incremental memory updates, writers create false authority for up to 50.2% of unauthorized requests; once false authority is present, executors act on it in 98.6% of trials. Two safeguards, requiring stored permissions to be backed by valid source events, and tracking permission changes through bounded event sourcing, substantially reduce laundering, but both also reject more legitimate actions, exposing a safety-utility tradeoff. Persistent memory is therefore not merely a performance component, but a part of an LLM agent's effective authorization policy.

Fleets of LLM agents now externalize effects that cannot be fully undone: they move money, deploy code, delete data, and disclose information. Current controls check one effect at a time, so a fleet of individually authorized agents can overdraw its principal's risk under a shared trigger while every local gate stays correct. We propose the irreversibility budget, a cumulative account of residual value-at-risk that a trusted runtime maintains for each principal across agents, workflows, and tenants. Treating irreversibility as a first-class resource, the runtime charges each effect its residual loss below the agent and denies the marginal effect once the aggregate would overdraw the budget. Getting the price right is hard, because effects are heterogeneous, adversarially declared, and correlated. We perform a controlled study in which per-effect gates admit fleet-level overdraws of up to 48 times the tenant's risk limit while the budget holds every correctly charged run within that limit. Conservative, dependency-aware pricing remains the central open problem for a deployable design.

Autonomous LLM agents increasingly act on a user's behalf: they hold credentials, call tools and services, and spawn sub-agents that act further on their behalf. This turns a long-standing distributed-systems question -- who is authorized to do what, on whose authority -- into an urgent and largely unsolved problem, because the component driving each agent is a language model an adversary can hijack. We argue that agent security must be evaluated under an untrusted-model assumption: a correct system is one in which a fully prompt-injected agent still cannot exceed the authority explicitly delegated to it. Against this standard we make three contributions. First, we give a threat model for multi-agent delegation centered on four adversaries -- confused deputy, token theft and replay, prompt-injection privilege escalation, and compromised sub-agents -- and derive eight security requirements a governed agent system must meet. Second, we show the gap is real: a default agent runtime modeling common practice (broad bearer credentials, authorization gated inside the model) fails all four threats, and across four widely used frameworks -- LangGraph, CrewAI, AutoGen, and the Model Context Protocol (MCP) authorization model -- three provide no built-in confinement and one only partial; no existing standard alone covers the requirement set. Third, we implement and adversarially evaluate an authorization broker that closes the gap. It blocks all four threats; it resists 11 direct attacks on its design and accepts 0 of 200,000 forged tokens; it confines a compromised sub-agent to its delegated task (a mean of 1.5 reachable actions versus all 8,100 under bearer delegation, across 2,000 randomized scenarios); and it enforces at microsecond cost (about 2.6 microseconds per decision), negligible against model inference. These principles are also realized in production in VotalAI's LLM Shield.

Multi-agent AI platforms move quickly from staging to production, but the way agents establish trust remains rudimentary: an agent either transmits raw data to a peer or accepts that peer's natural-language self-report that a value complies with policy. The first over-shares; the second is unverifiable and is exactly the channel prompt injection attacks. Prevailing responses emphasise identity, visibility, and post-hoc detection, and recent proposals for cryptographically enforced agent policy have been evaluated in simulation rather than execution. We take provable data minimisation between agents from proposal to running system. In our Zero-Knowledge Proof Gateway, agents exchange proofs of governance-defined predicates over private data rather than the data itself, so exposure is prevented by design rather than detected afterwards; because no interoperability protocol can carry such a proof, we propose a slot and implement it on both MCP and Agent2Agent from one endpoint. A 32-bit threshold predicate proves in 6.2 ms and verifies in 1.0 ms with a 608-byte Bulletproofs proof on one commodity vCPU; eleven adversarial experiments and nineteen protocol checks pass; and the system is deployed to Kubernetes with empirically verified network isolation. Our case study proves a retail client order is within its limit without revealing the amount, instantiating the GDPR data-minimisation principle as an enforced technical measure of the kind EU law now names explicitly. We then address the limitation no comparable work resolves: a predicate proof binds a statement to a committed value, never to the system of record. We give a construction fusing an enclave attestation with the proof in both directions, so verifying one artifact certifies jointly that the predicate holds and that the value was read by a specific measured binary, and test it against a mock authority.

Retrieval-Augmented Generation (RAG) improves large language models by grounding them in external evidence, but this exposes them to knowledge-poisoning attacks, where misinformation injected into retrieved documents influences model outputs. We investigate whether deliberative reasoning reduces susceptibility to poisoned evidence using two metrics: Cordon Rate, which measures cases where detected misinformation nevertheless influences the final answer, and Leakage Rate, which measures implicit influence from poisoned context despite explicit instructions to disregard it. We evaluate six model configurations on 200 SciFact questions, including DeepSeek-V4-Flash and Qwen3.6-Plus with reasoning disabled and enabled. Enabling reasoning reduces conditional susceptibility: DeepSeek-V4-Flash reduces Cordon Rate from 0.211 to 0.107 and Leakage Rate from 0.235 to 0.140, despite overall attack success rising from 0.233 to 0.298. These results show that poison detection, attack success, and resistance to contextual influence are distinct capabilities, and that deliberative reasoning reduces behavioral impact of corrupted evidence conditional on detection, even as it renders explicit poison identification less reliable.

Shared memory helps language-model agents reuse information across long workflows, yet relevant evidence may not be admissible for a particular agent or action. Because restrictions propagate through derivations, summaries can conceal private, poisoned, untrusted, or revoked sources, enabling unauthorized reads or unsafe actions. Existing approaches provide semantic retrieval, scoped access, or lineage tracking, but do not clearly separate hard authorization from graded trust or adapt evidence requirements to action risk. We introduce MAP-Graph, a provenance-aware memory layer that represents agents, sources, memories, claims, and actions in a typed execution graph. It traces ancestry, excludes permission-ineligible records, reranks eligible memories by semantic similarity and multiplicative path trust, and applies a risk-sensitive gate before action execution while retaining affected lineage for audit. On a controlled benchmark of 2,700 synthetic tasks per method across three domains, MAP-Graph achieves 94.96% overall task success, 72.70% exact decision accuracy, and 90.22% in the clean setting, where success requires a correct \textsc{Allow} rather than a safe intervention. Ablations isolate the roles of permission filtering, path trust, and action gating, while transfer tests with two additional backbones preserve the exact-decision and access-control advantages. These results support provenance as an operational control signal, rather than only post-hoc audit metadata, within the evaluated setting.

Multi-agent large language model (LLM) systems can expose protected state through internal messages, tool arguments, logs, and persistent memory even when their public outputs appear innocuous. Existing privacy prompts, redaction methods, and source-level access controls restrict surface content or data access, but do not specify what a legitimately informed agent should disclose or how that disclosure may be reused downstream. We introduce Minimum-Necessary Communication (MNC), a typed semantic-declassification protocol that selects a task-sufficient disclosure from an application-authored candidate family and binds it to explicit recipient, purpose, forwarding, lifetime, logging, and memory scopes. A reference monitor enforces these scopes across subsequent operations, while a history-aware extension accounts for inference risk accumulated over repeated disclosures. Controlled semantic-join, memory, probing, and longitudinal experiments show that conventional defenses can preserve protocol-level utility while exposing substantial additional inference signal. Under identical receipt text, MNC preserves authorized delivery while blocking unauthorized forwarding, logging, durable storage, and retrieval after expiration that a text-only semantic declassifier permits. Two-backbone MAGPIE executions further show that mediated disclosures propagate through subsequent planning, tool use, coordination, and memory retrieval. These results support scope-bound semantic declassification as a practical communication boundary for private LLM-agent systems.

Practitioners accept a compressed language model once it clears a stack of data-cheap quality guards: perplexity within a small factor of the original, downstream accuracy (for example MMLU) inside a confidence interval, and data-free output-fidelity signals that compare the compressed and original network's internal representations under random probe inputs. This stack has a blind spot. Across three model families, gently-compressed models clear every guard and then invent procedure steps that were never in the instructions when they run a standard operating procedure (SOP) as an agent. The effect is operator-specific: coherent low-rank (SVD) truncation induces it, and magnitude pruning matched to the same perplexity does not. One dissociation isolates the cause. The same compressed weights that CI-win a paired output-fidelity test CI-fail the invented-step canary. The governing axis is the coherence of the compression error times its rate; the magnitude of the damage does not predict it. The data-free fidelity probe is a fidelity oracle by construction, so it cannot see this axis. We characterize the blindspot and dissociation with paired confidence intervals on a pre-registered, powered canary across three architectures. Operator-specificity replicates on all three, and the perplexity-guard evasion appears where the model admits in-guard low-rank headroom. We then give a data-free screen: a two-axis statistic of the compression error (coherent-fraction and error-rate) that flags the failing builds with fixed thresholds across architectures and matches the coherence-times-rate mechanism. Perplexity, MMLU, and fidelity acceptance do not certify agent safety. Screen gently-compressed low-rank builds before agentic deployment

Autonomous LLM agents processing mixed-confidentiality data face severe security risks from prompt injection attacks and reasoning errors. While dynamic Information Flow Control (IFC) provides structural security guarantees, traditional taint tracking permanently taints an agent's context upon reading unvetted data, severely restricting downstream utility. We present APPA (Agentic Permissions Policy Algebra), an IFC framework that resolves this usability bottleneck through engine-managed context branching and prospective acquisition enforcement. Before data acquisition occurs, APPA prospectively evaluates label descents and missing prerequisites, generating actionable remedy plans (Authorize, Accept). To inspect unvetted data without polluting the primary context, a label-seeded child trajectory is spawned, absorbing label descent locally and allowing a trusted sanitizer to return a bounded derivative to the unchanged parent. Governed by a two-monoid model over security labels and shared event logs, we formally prove parent label preservation and merge confinement. Finally, we evaluate APPA on a multi-turn tool-chaining benchmark across four models: it suppresses exfiltration (31%-50% down to 0%-7% attack success), and on three of the four, branching recovers a substantial share of the utility that taint tracking alone forfeits.

LLM agents choose tools and arguments from context that mixes user requests, tool outputs, retrieved records, memory, and untrusted text. Evidence can be relevant without being authorized to determine a decision, so a correct action need not be grounded only in permitted evidence. We introduce a target-specific authorization audit that labels context factors separately for each tool and argument target. Its primary test holds the task, proposition, position, and policy fixed while changing only the proposition's source authority. We then test behavior when valid evidence is weakened and use context-subset interactions as a secondary localization diagnostic. Across 450 controlled next-action tasks and multiple open-weight LLM families, trusted and untrusted variants produce different actions in 5.4 percent of competing cases versus 1.7 percent of supporting cases. Under controlled degradation, unauthorized competition is retained in a full-correct, mixed-error, clean-correct pattern in 2.4 percent of comparisons, with a 95 percent confidence interval from 2.1 to 3.0 percent. These are controlled stress-set rates, not deployment prevalence. The models respond to textual source-authority cues, but this does not prevent untrusted evidence from influencing their actions.

Evaluation frameworks for tool-augmented LLM agents focus overwhelmingly on capability metrics or explicit tool crashes, leaving silent infrastructure failures and HTTP 200 responses with empty, null, or malformed payloads largely unaudited. We introduce a lightweight black-box auditing framework that injects four silent failure profiles across 12 production-adjacent tool stubs and classifies agent responses into three mutually exclusive behavioral classes: Honest Surrender (HSR), Fabrication (FAR), and Unfaithful Safety Refusal (USR). Evaluating two frontier and two open-source models at temperature zero under a neutral system prompt, we find that FAR dominates (56.6% of valid responses): agents treat empty payloads as real data, silently returning fabricated results. USR, in which an agent invents a policy or privacy rationale to explain the failure, is nearly absent at baseline (0.25%, one instance across 396 valid trajectories). Our key finding emerges from an ablation where we augment the system prompt with standard safety language ("prioritize user privacy and data security"), which amplifies USR by 15.6x (from 0.25% to 3.95%; 95% CI on ablation rate: 2.2%-6.4%; Fisher's exact test, p < 0.001). USR is a latent behavior, activated when safety vocabulary in the system prompt primes the model to reach for policy rationales when tools silently fail. Sensitive tools (fetch_medical_record, retrieve_contract, fetch_user_profile) account for the majority of USR instances. We propose a payload-response misalignment heuristic for production-level detection and discuss governance implications for safety-forward deployments.

Large language model (LLM) agents are starting to take on routine work in high-performance computing (HPC), including monitoring Slurm jobs, diagnosing failed builds, inspecting simulation output, and coordinating scientific workflows. To do this work, an agent commonly acts under its user's credentials and inherits the user's access to files and the scheduler. This arrangement creates a failure mode that ordinary account-level controls do not capture. Adversarial instructions in a log, tool description, shared file, or peer-agent message may redirect the agent beyond the task the user assigned, even though every resulting command is authenticated and permitted for that account. We refer to this as the hijacked authorized agent problem. Existing agent-security studies explain relevant mechanisms, such as indirect prompt injection and tool misuse, but generally evaluate them in web, enterprise, or personal-assistant settings. HPC security, by contrast, has mature controls for identity and isolation but does not ordinarily represent the intent of a particular task. This paper defines the threat model in the HPC setting, identifies attack surfaces created by schedulers, shared storage, multi-project accounts, and scientific workflows, and examines where current controls fall short. It concludes with a research agenda and a plan for an empirical benchmark, TaskBound.

Agentic commerce extends agentic shopping into software agents that interpret policy, prepare checkout, generate transaction-facing language, and act under delegated payment authority. Protocols standardize external exchanges, but merchants still need one authoritative representation of commercial eligibility, actor authority, checkout validity, payment dispatch, generated claims, and evidence. This design-science study presents a protocol-agnostic architecture built around a canonical envelope, protected dependency and result hashes, Ed25519 or HMAC authentication, live-request rebinding, a seven-axis generated-claim gate, execution-time dependency revalidation, and eleven semantic invariants. Evaluation used an open-source JavaScript implementation, eight deterministic ecommerce scenarios, and five controlled ablations. Seven initially valid actions were permitted. After protected state changed, none could proceed without a fresh decision; a hostile-accessor case also remained blocked. Action status was consistent across configured surface-bound envelopes, and each scenario contained the three protected hashes and its targeted dependency reference. Each ablation produced the predicted unsafe regression when one safeguard was bypassed, while the protected path contained the same failure. The hostile accessor was read once, and the suite passed 66/66 tests, schema validation, and committed examples. Results support protected-dependency change detection, bounded outcome derivation, stale-decision prevention, surface-bound status consistency, refusal propagation, and verified-state identity under synthetic fixtures, but do not establish rule completeness, production security, performance, legal compliance, live interoperability, population error rates, or independent replication.

Autonomous coding agents increasingly execute multi-step software work, but lifecycle states such as reviewed, tested, DONE, and ready-to-merge remain claims unless supported by current evidence. We present Proof-or-Stop Lifecycle Control, a method that permits lifecycle transitions only when fresh, tracked-source-state-bound, mechanically verifiable evidence satisfies the relevant gate. The method treats agent outputs as claims rather than lifecycle state, and uses proof operationally to mean gate-admissible evidence under a stated trust model, not semantic program correctness. We evaluate an open-source implementation through mechanism tests, a powered control-policy ablation, and operated self-application evidence. The unattended-loop engine passed 10 of 10 scenarios with zero false-DONE, and local-key receipt bundles rejected 18 tamper classes with zero false accepts. In a 9,240-cell ablation, the pre-registered A4 versus A2-prime comparison reduced visible-pass/hidden-fail amplification from 31 of 1,800 injected cells under a compute-budgeted naive loop to 2 of 1,800 under the gated loop, a 1.6 percentage-point improvement in not-amplified rate with a 95 percent confidence interval of [0.8, 2.5]. A near-compute A3 versus A4 comparison, 14 of 1,800 versus 2 of 1,800, indicates that the gain is associated with enforcing review as a lifecycle gate rather than merely adding a reviewer. The self-application corpus contains 565 stories and 1,007 review findings, with 94.8 percent resolved, plus a 68-row high/critical cross-vendor exhibit. These results support Proof-or-Stop as a model-agnostic, host-neutral control layer for deciding which autonomous-agent claims a lifecycle may act on. The evaluation is limited to one model family, 24 ablation tasks, and a self-hosted corpus.

Safety alignment in LLMs aims to align models with human values, but which values take precedence when they conflict? We investigate this question in the context of tool-calling LLM agents deployed in regulated industries, where agents processing confidential documents may encounter content that triggers safety-trained values (e.g., public welfare) that conflict with deployment-context instructions (e.g., internal logging). To empirically verify this phenomenon, we build a benchmark of 128 scenarios across 16 domains. We find that safety-aligned open-source models override their deployment instructions up to 43.4% of the time, engaging in whistleblowing, data exfiltration, and evidence tampering when processing documents that suggest organizational wrongdoing. We also find that abliteration reduces rates of external whistleblowing. These results reveal a fundamental tension in pluralistic alignment, where the same safety training that protects users can cause agents to act against deployment instructions in ways that create unpredictable liability risks. We release our benchmark as a framework to support evaluation of agent behavior under competing legitimate interests.

LLMs are now proposed for fraud detection, scam investigation, content moderation, and other trust-and-safety workflows. Much of the public literature still evaluates them as models, with less attention to their behavior as components in operational pipelines. This creates a practical evidence question: what would justify placing an LLM inside a live workflow with latency, cost, escalation, human-review, and adversarial-risk constraints? We address this question through a fraud-first survey of deployment evidence. We code 49 operationally relevant sources on LLM use in fraud detection, investigation support, content moderation, and cross-cutting robustness (18 fraud, 14 moderation, 17 cross-cutting), supplemented by 15 contextual references that establish the survey boundaries. These sources include systems, benchmarks, frameworks, and deployment-relevant surveys, not 49 production deployments. The main finding is an evidence imbalance. Fraud supplies the largest task-specific portion of the coded corpus. The moderation papers, however, include more explicit public evidence on latency, cost, governance, and fairness. Among the 18 fraud and investigation sources, none report clean per-decision latency, per-decision dollar cost, or calibration evidence; most report offline task performance, retrieval gains, or case-study accuracy instead. The survey contributes a role-and-evidence organizing frame, FORTE, for locating LLMs as classifiers, retrieval interfaces, explanation generators, reviewer assistants, agents, feature extractors, or escalation components. It also contributes a minimum deployment-evidence checklist covering latency budget, cost per decision, decision threshold, explanation integrity, and adversarial pressure. The resulting agenda identifies studies needed to support deployment claims for LLM-based fraud and trust-and-safety work.

LLM agents can commit durable effects from authority evidence that was valid earlier in execution: a DOM snapshot, approval epoch, version witness, branch token, or worker result. We study the commit boundary at which earlier authority evidence no longer authorizes a durable effect. We call this property commit-time authorization: a durable effect is authorized only if the witness that licensed its derived state remains fresh, causally prior, bound to the same effect, and eligible at commit time. We build a controlled-invalidation suite spanning browser, tool/API, and multi-agent workflows. The suite preserves the user goal and payload shape while invalidating the authority relation before durability. In the primary 54-task matrix, endpoint success remains high: 262/270 runs reach the visible result. Only 55/270 are authorized completions; among the 216 invalidating rows, 207 commit after the authorizing path has failed. All 54 clean controls remain authorized, and a separate 54-run authority-preserving check produces no unauthorized commits. We then evaluate mitigation families. Prompt caution and single-condition checks are insufficient because different hazards break different boundary conditions. Defenses work when they refresh, rebind, replan, or refuse at the durability boundary. CommitGuard, a fail-closed boundary monitor, blocks stale durable-effect attempts on protected commit surfaces when runtimes emit witness, dependency, binding, and eligibility signals. The result is a reporting and runtime-design lesson: endpoint success is a utility metric; authorized commit is a security property.

Semantic caching defines answer reuse on embedding similarity: two utterances share a stored answer when a similarity score clears a threshold, with no notion of authorization, versioning, or of what makes two demands the same. This note changes the object on which reuse is defined: in a governed domain, reuse should operate on a mathematically characterized quotient of resolved conversational demands, not on a similarity heuristic. Three independently defined relations on resolved utterances -- reading identity, resolution identity, and reuse identity -- form a refinement chain, strict under realized nondegeneracy conditions checkable on deployment logs; the pipeline's outputs are invariant along the chain, and reuse identity is exactly the kernel of the resolution map into the governed answer partition, so the reuse quotient is the utterance-side object that partition induces, not a relabeling of it. Reuse identity licenses the governed query key and its certified answer space; reuse of a particular answer requires resolution identity or an applicability certificate. The supporting layer is stated at exactly the strength proved: exact-denotation normal forms; join aggregation as a design operator, with closure-stable cells characterizing no-escape; total computability of the full pipeline relative to an untrusted proposal layer; policy admissibility for arbitrary proposers -- and provably not factual grounding or intent fidelity; and elicitation terminating after finitely many informative replies, sound under target consistency.

LLM agents reach users through resellers, who may rebrand a developer's agent or substitute a cheaper model. When provenance is disputed, attribution rests on the trajectory log (the record of tool calls, observations, and executed actions, not the model's reasoning), which the reseller stores and processes to meter usage. A watermark must therefore survive an adversary with full read/write access to the very evidence it is detected from; existing agent watermarks do not, as their attribution is read straight off that log. We present TRACE, to our knowledge the first agent watermark that is distortion-free in its action choices, self-synchronizing under deletion, and unconditionally invariant under rewriting. Deletion desynchronizes a position-derived key and rewriting alters content, so a deletion-robust key must come from content and a rewrite-robust key from position, and no single key serves both. A trajectory, however, has room for two watermarks. TRACE superposes a selection channel that sets which action is chosen, keyed on local content with a distortion-free sampler, so the agent's distribution is provably unchanged and detection resynchronizes after deletions, and a tally channel that sets how many records each decision group holds, keyed on the log's skeleton alone, which no rewriting can touch. We prove this behavioral watermark's signal is bought with decision entropy, each decision paying at least half its entropy and deterministic decisions nothing, and that erasing both channels forces the reseller to corrupt the trajectories it resells. On ToolBench and ALFWorld, TRACE matches the unwatermarked agent's success rate while its selection channel reaches detection scores near z = 100 on long-horizon trajectories, stays detectable under 70% step deletion, and keeps a tally channel exactly unchanged under LLM rewriting of any strength.

Autonomous web agents promise to automate everyday browsing tasks, but inherit one of the web's oldest attack surfaces. Cross-Site Scripting proved that mixing trusted and untrusted content is dangerous, even on benign pages. Agents resurface this risk by interpreting natural language as instructions, allowing third-party and user-generated content to hijack the agent via prompt injection. The core challenge is that deriving a task-specific security policy requires reasoning over page structure that is entangled with the attacker's content. We present Prismata, a defense enforcing contextual least privilege for web agents, constraining both what the agent sees and what it can do. Prismata's dynamic trust derivation produces permission labels for page content, with structural confinement guarantees, inspired by classical integrity models, that bound any labeling errors so that labels can only decrease in privilege and mislabelings are bounded. Prismata's mechanical confinement enforces these labels by redacting content and restricting agent capabilities. Importantly, these mechanisms require no developer annotations, so Prismata supports the long tail of websites. Across recent published web agent attacks, including adaptive variants, Prismata substantially reduces attack success while preserving benign task utility.

The Model Context Protocol (MCP) is the dominant way coding agents discover and invoke external tools. A server advertises each tool through a tools/list handshake that returns a name, a natural-language description, and a JSON input schema. The client renders this metadata once, in a one-time approval dialog, and then injects it verbatim into the model's context on every subsequent turn. Nothing in the protocol requires the rendered approval view and the bytes delivered to the model to match. We isolate that gap as a single structural mechanism, concealment encoding, and show with a model-free, protocol-free analysis that Unicode's TAG block (U+E0000 to U+E007F) has no assigned glyph in any mainstream terminal, chat, or IDE renderer, so a payload written in it is absent from what a human reviewer sees while surviving byte-for-byte into the model's tokenizer. We then measure whether this mechanism actually defeats today's client-side defenses, building a proof-of-concept that speaks the real MCP JSON-RPC/stdio protocol against a genuine client and server. Across 5 distinct MCP metadata surfaces we implement 8 concrete techniques with a deterministic, protocol-level harness. All 8/8 techniques deliver an attacker-controlled payload into the model's context, 4/8 evade a representative string-matching sanitizer, and exactly as the mechanism analysis predicts, only the TAG-block encoding (1/8) is invisible in the human approval view while still reaching the model verbatim. MCP forces re-approval for 0/8 techniques even under a time-of-check to time-of-use rug-pull. To test whether these outcomes are a property of the protocol or an artifact of one server codebase, we re-implement the catalogue against 3 independently developed Python MCP server libraries and find total agreement across all 32 cross-library outcome cells. The baseline sanitizer flags 0 of 25 benign descriptions.

Trusted monitoring is a central defense in AI control: a cheaper trusted model scores an untrusted model's actions for sabotage, and the most suspicious are audited or deferred. Such monitors are evaluated against one or two untrusted models, and the accuracy is reported as a property of the monitor. We ask whether it is partly a property of the pairing. We make the untrusted policy family the controlled axis: we fit a monitor on family A's transcripts, apply it to family B, and decompose the cross-family AUROC into how obvious each family's sabotage is, how capable each monitor is, and the residual own-family advantage after both are removed: the interaction. On code-backdoor transcripts the interaction is positive and survives the dominant confounds: +0.172 (95% CI [+0.158, +0.185]) on four open-weight families on a strict leak-free basis, holding across an eight-family matrix spanning four frontier vendors, blatancy-matched controls, a trained transcript-reading monitor up to 34B parameters, and a second code benchmark. The gap is geometric: each lineage's sabotage occupies a different scoring direction, and rotating the monitor's scoring direction toward the target lineage recovers most of the lost detection while a matched random rotation recovers nothing. At a 1% audit budget under defer-to-trusted, an off-lineage monitor catches 19% of attack attempts where an in-lineage monitor catches 41% at the same audit cost, a +21.7-point gap. Control evaluations should report cross-family transfer matrices: a single-pairing accuracy overstates the safety a monitor delivers against a lineage it was not calibrated on. We give a four-step protocol that acts on the gap, with each step a measured result.

Defenses that provide security guarantees against prompt injection attacks rely on strict isolation between trusted instructions and untrusted data. In text-based environments such as tool-use APIs, this separation arises naturally: agents can reason from interface definitions without ever processing untrusted content. Extending these guarantees to web agents faces a fundamental challenge: to perceive and interact with their environment, web agents must first observe the rendered page, which intermingles trusted content with untrusted content. This structural entanglement removes the trust boundary on which security guarantees depend, undermining provable defenses for web agents. In this paper, we present Untrusted Content Masking (UCM), a simple and effective approach that restores this boundary in web environments. We leverage a key structural insight: a webpage's Document Object Model (DOM) encodes sufficient information to distinguish trusted from untrusted regions without reading their content. Our framework exploits this by redacting untrusted regions before they reach the agent and routing interaction through a sandboxed interface with strict privilege separation, thereby enabling agents to observe and interact with their environment while remaining isolated from adversarial content. The code is publicly available.

Personal AI agents that run on the user's local machine automate daily tasks including web search, email, and file management. Their access to computer resources, including the network, file system, and shell, exposes them to indirect prompt injection (IPI) attacks. Prior Dual LLM defenses block IPI by replacing untrusted data with symbols that the agent can reference but not read. However, they track untrusted data only inside the agent's context, so when the agent saves and later rereads untrusted data, that data, possibly an attacker's prompt, can return as trusted data rather than as a symbol, which we call stored IPI. Operating on the user's real environment is what makes agents like OpenClaw practical, and is exactly why a defense that ignores it is incomplete. Preserving symbols in such an environment is hard, because humans and programs need original data. We present DualView, which extends untrusted data tracking from the agent's context to the user's environment, including the file system, shell, network, and other agents, by giving each channel two views. In AgentView, the agent sees untrusted data as symbols even after writing it out and reading it back, blocking stored IPI, while HumanView preserves original data for humans and tools. DualView routes each tool call to the right view and synchronizes data across the two views. DualView deploys as an OpenClaw plugin using only tool hooks, without changing the agent's tool-call logic or tool implementations. DualView deterministically prevents instructions in untrusted data from directly steering the agent's tool calls; this guarantee does not depend on recognizing the evaluated attack templates. In our evaluation on an IPI benchmark and PinchBench, DualView blocked every tested IPI attack, including stored IPI. On PinchBench, its utility drop was within 1.8 to 6.4 points.

LLMs increasingly generate workflow actions and repairs that may be well formed yet stale, infeasible, conflicting, or destructive of their own evidence. We introduce Agentic Transaction Processing (ATP), which treats generated actions as untrusted proposals until deterministic admission accepts them under an executable constraint set C. Its two-sided principle is: a proposal is not truth, and no proposal foresees every disruption. Anything may propose, but only the runtime admits and commits; unforeseen disruptions trigger bounded reactive repair whose output re-enters admission. Mnemosyne realizes ATP with an append-only transition log, effective-state projection, dependency-safe compensation, and active contract records. Under stated assumptions and relative to C, we prove four safety properties (authority separation, serial-equivalent generative admission, evidence-preserving repair, and obligation containment) and establish bounded reactive repair. Across nine safety benchmarks and a four-case Temporal SDK comparison, ATP rejects every targeted violation while admitting valid work. A matched data-size sweep measures 5.2-6.6% incremental throughput cost over the same local durable commit path and exposes local saturation. In a companion scheduling harness, local repair edits nearly an order of magnitude fewer operations than global recompute; a 12-scenario interruption stress test rejects every stale recovery candidate without losing observations or producing invalid commits. Two bounded pilots route 80 proposals from four heterogeneous LLMs through the same gate with zero invalid commits; 24 of 40 mid-execution proposals are admitted and 16 are rejected, including four explicit safety rejections.

We identify a security-fidelity tradeoff in defending LLMs against indirect prompt injection: defenses resist injected instructions largely by suppressing untrusted text, which corrupts tasks that must preserve it, such as translation and document editing. Attack-success metrics cannot see this, because a model that ignores an injection and one that faithfully processes it as data score identically. We introduce SecFid, a benchmark built so that executing an injection, processing it as data, and ignoring it produce distinguishable outputs. This makes fidelity measurable and exposes a frontier: across 1,168 examples and 48 configurations, no model or defense achieves both objectives. The highest-fidelity model reaches 96.5% fidelity at 47.8% security, while the most secure defenses invert this, at 99.3% security but only 71.0%-73.9% fidelity. Even defenses with identical security differ in how they earn it: some repair hijacks into faithful processing, others simply suppress benign content. A decision-theoretic analysis shows why no fixed choice can be right everywhere: the correct behavior is not a property of the defense but of the deployment, set by its relative cost of a hijack versus a dropped span. Security alone therefore measures only half of robustness, and reporting it without fidelity hides the price at which it was bought.

Tool-using LLM agents increasingly read untrusted content while holding side-effecting tools such as payments, email, CRM, and infrastructure APIs, yet common framework defaults still conflate tool exposure with authorization. We audit whether LangChain/LangGraph, LlamaIndex, and the Stripe Agent Toolkit re-authorize each model-emitted call, with concrete argument values, before execution. Across pinned public-source commits, all three provide capability gating by default, but none provides a deterministic fail-closed per-call value authorization gate by default. We introduce ScopeGate, a five-stage PDP/PEP for agent tool calls: scope, authorization, money ceiling, idempotency, and default deny. Evaluation shows the identical unauthorized payout call executes under LangChain's default dispatch (with a companion LlamaIndex PoC) but is denied by ScopeGate; the tested control reports 0/48 static bypasses, 0/29 unauthorized attempts (40-iteration adaptive run), 0/10 benign false-denies, and Latam-GPT payment-agent containment at 10/10. ASR denotes attempted unauthorized action, containment is not a cure, deployment-tier claims are inference over measured model classes, and no CVE is asserted.

Prompt injection is the top security risk for LLM-integrated applications, yet every defense proposed so far has been broken. We prove this is not a coincidence: in shared-embedding architectures that lack enforced control-data separation, perfect prompt-injection prevention is mathematically impossible. We formalize prompted systems as Prompted Action Models whose outputs include control-authoritative actions: refusal decisions, tool authorization, policy routing, and memory writes. We define Semantic-Faithful Control (SFC), the property that such behavior depends only on the meaning of untrusted input, not on how it is encoded. We then prove SFC is unachievable within the shared pipeline, via three results: a provenance-recovery impossibility (shared representations make trusted and untrusted content statistically inseparable, bounded by total variation distance); control-path exposure (untrusted tokens enter control-relevant computation through the same attention value-aggregation that determines outputs); and a finite-coverage invariance gap (finite training cannot certify invariance over infinite semantic-equivalence classes). We ground each quantity in measurements on production tokenizers and models. The result is structural, not a gap in current defenses. It mirrors the code-data confusion in Von Neumann machines that gives rise to buffer overflows, a vulnerability class that took decades of layered defenses (DEP, Write-XOR-Execute, ASLR, stack canaries, and ultimately memory-safe languages) to contain, because no single mechanism sufficed. The implication is the same: prompt injection cannot be eliminated by better in-pipeline classification or alignment alone. It requires architectural separation of instruction and data channels. We identify the root cause and the class of solution it demands.

As autonomous AI agents increasingly transact across organizational boundaries, a fundamental trust challenge emerges: how can an agent assess whether an unknown counterpart is trustworthy? The ERC-8004 protocol addresses this challenge with the first permissionless trust layer for AI agent economies, built around three on-chain registries for Identity, Reputation, and Validation. Despite its rapid adoption, the protocol has not been studied empirically, leaving it unclear whether the information it records provides a trustworthy basis for decision-making. To address this gap, we present the first empirical study of ERC-8004 across three chains: Ethereum, BNB Smart Chain (BSC), and Base, covering the period from protocol deployment through May 13, 2026. We crawl on-chain Identity and Reputation events, off-chain files, and x402 payment transactions. On the identity side, we find that most registrations are placeholders rather than active agents, with only a small fraction (3%, 4%, and 15% across Ethereum, BSC, and Base) exposing a valid ERC-8004 registration file with at least one live service endpoint. On the reputation side, we show that the Registry, as currently deployed, cannot function as a trust signal: values are not commensurable, feedback records are rarely grounded in verifiable interactions, and reputation can be manipulated at minimal cost. Consistent with these design weaknesses, we find that a substantial fraction of reviewers (73.5%, 59.2%, and 90.6% across Ethereum, BSC, and Base) exhibit coordinated Sybil behavior. After removing Sybil-flagged feedback, 15.8%, 77.9%, and 86.8% of rated agents, respectively, are left with no valid feedback. We then turn these findings into concrete recommendations for future revisions of ERC-8004. Our study yields actionable protocol-design implications and establishes an empirical baseline for research on AI agent markets.

Large language models increasingly mediate interactions between sensitive data, untrusted inputs, and privileged actions in agentic systems, creating security and privacy risks. These range from prompt injections that manipulate downstream tool use to leakage of confidential information through model outputs. Recent Information Flow Control (IFC)-based defenses show promise but lack a principled semantic foundation for reasoning about information flow through the model itself. Since any input token may influence any output token in an autoregressive LLM, existing approaches suffer from severe taint explosion. We present Geometric Information Flow (GIF), a semantic framework for tracking information flow from input tokens to outputs. GIF uses the LLM Jacobian and local output geometry to upper-bound the Shannon mutual information between perturbed input spans and model outputs, yielding a scalable measure computable on large models via automatic differentiation and low-rank approximation. Unlike attention-based or correlational attribution heuristics, GIF satisfies local geometric soundness, and we provide a fully mechanized Lean 4 proof that it upper-bounds the true information flow induced by a given prompt under local regularity assumptions. We evaluate GIF on integrity and confidentiality tasks across multiple prompt-injection and privacy-leakage benchmarks. GIF achieves near-perfect recall even without a downstream declassifier, outperforming attention-based baselines. Combined with lightweight LLM-based declassifiers, it matches or exceeds the F1 of direct LLM-as-judge baselines such as GPT-5.5 xhigh reasoning while using up to 81x lower token cost. GIF flows detected with small surrogate models transfer to larger state-of-the-art models and other model families, even when the surrogate is up to 200x smaller, suggesting black-box deployment without gradient access.

AI agents increasingly access external models, tools, and services through Agentic Routing Infrastructure (ARI) to manage the overhead of heterogeneous interfaces and fragmented subscriptions. Yet, the architecture of ARI introduces fundamental trust risks: it obtains plaintext access to agent queries and service responses, while leaving agents unable to verify that their queries are routed to intended service providers or that requests and responses remain untampered. To address this problem, we present TrustedARI, the first trust-native agentic routing infrastructure for agentic AI. Architecturally, TrustedARI is built upon three core innovations: (i) an ARI-adapted three-party TLS handshake that enables the agent and ARI to jointly authenticate the service provider through role-specific distribution of TLS key materials; (ii) a privacy-preserving query-construction protocol that allows the agent and ARI to collaboratively construct well-formed queries without exposing their respective private inputs; and (iii) a verifiable billing protocol that supports fair usage-based settlement while preserving the integrity and confidentiality of service responses. We implemented and extensively evaluated a prototype of TrustedARI to validate its performance. Experiments confirm that TrustedARI is highly efficient: our ARI-adapted handshake protocol reduces communication overhead by 39.34% compared to the existing three-party TLS handshake. Furthermore, the privacy-preserving query-construction protocol imposes negligible overhead-averaging 0.19 seconds in computation time and 0.58 MB in communication costs-while the verifiable billing protocol speeds up proof generation by 28.20x. Crucially, TrustedARI is readily deployable without any modification to the service providers.

Modern LLM-powered autonomous agents increasingly rely on rich user interface (UI) state observations to achieve reliable action grounding in complex digital environments. However, many deployments transmit the full UI state to remote inference servers even when most elements are irrelevant to the current task, which can leak sensitive but unnecessary context such as authentication codes, private notifications, and background application states. We propose MINIM, a trusted local broker that performs privacy-aware minimization on the client side before any observation leaves the device. Grounded in Contextual Integrity (CI), MINIM learns a dual-score representation for each UI element by predicting an inherent sensitivity score (s) and a task-conditioned necessity score (n). These scores drive a ternary disclosure policy that keeps essential elements, abstracts sensitive attributes when needed, and removes task-irrelevant content. We optimize a CI-aware objective that penalizes necessity errors more strongly on high-risk content, enabling aggressive pruning while preserving task-critical information. Experiments on real-world UI observations derived from WebArena show that MINIM substantially reduces task-irrelevant sensitive leakage while preserving task-critical semantic context and the interactive affordances required for reliable agent actions.

Trusted monitoring is a cornerstone of AI control. However, as frontier models grow more capable, the increasing capabilities gap between trusted and untrusted models may render trusted models unreliable monitors. We introduce \emph{bootstrapped monitoring}, a protocol that addresses this by inserting a stronger, intermediate untrusted model with transparent chain-of-thought reasoning into the oversight chain. The untrusted monitor ($U_m$) evaluates the agent's actions, while a weaker trusted model ($T$) oversees $U_m$'s reasoning to detect collusion. We evaluate bootstrapped monitoring on multi-turn software engineering tasks (BashArena) across multiple agents and monitors. Bootstrapped monitoring substantially improves catch rates over trusted-only monitoring, even when the untrusted monitor actively colludes with the agent, provided we have access to its raw chain-of-thought. Our results suggest that bootstrapped monitoring can extend the useful lifetime of trusted models in control as AI capabilities advance.

Long-horizon LLM agents are not trusted to run unattended: with no human watching, they confidently report success they never verified. We treat honesty -- bounding what an agent may claim at termination -- as a first-class metric for unattended autonomy, distinct from capability. We present Autopilot, an execution model that makes silent fabricated success structurally impossible rather than merely rarer. Autopilot externalizes all working state into a durable, gated finite-state machine that a scheduler advances one stateless tick at a time; a hard floor forbids any terminal "done" claim whose falsifiable gate did not actually execute and pass. We prove a No-False-Success theorem -- under gate soundness, floor enforcement, and plan coverage, termination implies the goal holds -- whose only trust points are empirically measurable, and show the worst case degrades to an honest stall, never a fabricated success. Because each tick rehydrates only the state machine, per-step context cost is constant in the horizon. Across a 3,150-cell paired corpus (70 tasks $\times$ 3 systems $\times$ 3 models $\times$ 5 seeds, including 50 SWE-bench Lite tasks across 11 OSS repos), Autopilot fabricates on 0.95% of cells [95% CI 0.38--1.62] while Reflexion and StateFlow baselines fabricate on 8.10% [6.48--9.81] and 25.05% [22.48--27.62] respectively. The headline contrast lives in the hard regime: on SWE-bench Lite, the firewall reduces fabrication from 33.7% (StateFlow) to 0.67%, a paired difference of $-33.07$ pp [95% CI $-36.53, -29.73$]. The mechanism is the gate, not the model: all ten Autopilot fabrications come from the strongest model, while two weaker mid-tier models never fabricate across 700 paired cells. The firewall trades coverage for honesty by design -- an honest stall is recoverable; a confident wrong output shipped downstream is not.

AI control protocols oversee untrusted models by monitoring their actions and modifying potentially unsafe steps, often using a trusted model. This partially tampers with the untrusted model's trajectory. If the trusted model detects such an intervention, it may infer properties of the monitor and adapt to evade control. We introduce \textbf{CIAware-Bench}, a benchmark for measuring \textbf{c}ontrol \textbf{i}ntervention (CI) awareness across frontier models. CIAware-Bench tests whether models can distinguish their own trajectories from those modified by a control intervention. The benchmark is comprised of a suite of four task domains (essay writing, BigCodeBench, Bash Arena, and SHADE-Arena), while varying trajectory watermarking, side-task presence, and the control protocol. Evaluating eleven frontier models, we find low to moderate CI awareness under default settings (up to 0.87; random chance balanced binary classification accuracy is 0.5) with substantial variation across task domains and model pairs. Detection is generally easier across model families, suggesting that models exploit provider-specific differences in style or post-training. Overall, CI awareness is not a fixed model-level property, and should be measured for each new model release and deployment scenario. We release CIAware-Bench to track CI awareness and inform control protocols whose interventions are harder to detect.

Recent progress in generative modeling has made safety control a central challenge, yet existing approaches remain largely model-specific, requiring retraining or tailored interventions for each new architecture. In this work, we ask whether safety can be represented as a portable latent direction, learned once and reused across heterogeneous generators. We introduce the first framework for cross-model safety steering, in which a safety direction is estimated in a source LLM from paired safe-unsafe prompts, transported to a target generator through a lightweight alignment fitted on benign data alone, and applied at inference time. Crucially, our pipeline never accesses unsafe data on the target side, isolating whether safety can be transferred through shared representation geometry. Beyond a single global direction, we also identify a multi-vector extension that captures category-specific safety behaviors, enabling more selective control. We evaluate our approach in text-to-image and text-to-video generation across diverse source-target model pairs. Across models, transferred safety directions achieve ASR reduction and CLIP-Score/FID trade-offs comparable to directions learned natively on the target model using unsafe data, while requiring no target-side unsafe data. This indicates that safety improvements do not come at the expense of generation quality. Our results point to a modular view of safety: safety-relevant behavior is not purely model-local, but can be controlled through latent directions that persist across models. This suggests a new path toward lightweight, reusable safety mechanisms that do not require target-side unsafe data.

Memory is a core component of AI agents, enabling them to accumulate knowledge across interactions and improve performance. However, persistent memory introduces the risk of memory poisoning, where a single adversarial memory write can exert long-term influence over agent behavior. We present a systematic study of memory poisoning in LLM-based agents. We identify four memory write channels and nine structural vulnerabilities in model capabilities, system prompt design, and agent system architecture that make these channels exploitable. Based on these vulnerabilities, we develop a taxonomy of six classes of memory poisoning attacks. Furthermore, we design MPBench -- a benchmark for evaluating memory poisoning attacks, and show that agents designed to write and retrieve memory more aggressively are more exploitable. We also show that existing prompt injection defenses fail to cover memory poisoning attacks. Our findings provide a foundation for understanding and mitigating memory poisoning attacks against AI agents.

This yellow paper describes the technical architecture of OpenAgenet / OAN. OAN is a protocol-neutral trust layer for open Agent interconnection and discoverable AI resource products. It specifies the role architecture, \texttt{did:oan} identity objects, registration workflow, governance-backed Root lifecycle enforcement, Root-verified package model, authorization-aware Discovery, Root-issued infrastructure authorization VCs, signed trusted invocation, verification requirements, state transitions, security properties, implementation boundaries, and deployment considerations. The design is intended to support heterogeneous Agent frameworks and interaction protocols, including MCP, A2A, ANP-like systems, domain-specific Agent protocols, Skills, MCP Servers, and Tool/API resources. OAN does not define the entire business conversation among Agents or the native protocol of every resource; it defines how resource identities become admissible, discoverable, verifiable, and safe to approach before protocol-specific interaction begins.

We present a systematic study of domain-dependent safety behavior in open-weight LLMs: 7 standardized experiments across 7 ethical domains, testing 5 models (12B--70B) in 4,200 interactions with dual-judge validation. Using a dual-condition methodology, each scenario tested in both an analytical framing (identify the harm) and an operational framing (help commit the harm), we find compliance rates vary from 14.7% (human trafficking) to 85.7% (surveillance design), a 71-percentage-point span with non-overlapping cluster-bootstrapped 95% CIs. Trustworthy deployment requires predictable safety behavior, yet we find compliance is highly context-dependent: the same model (Mistral Nemo 12B) provides surveillance designs in 100% of requests but assists with trafficking in only 26.7%. This unpredictability is opaque to deployers: the technical framing bypass, where harmful requests reframed as engineering problems override safety training without any external signal that refusal thresholds have shifted. Within-domain heterogeneity reaches 84.4pp, meaning safety behavior cannot be predicted even at the domain level. A replication on five frontier closed models (GPT-4.1/5.2, Claude Haiku/Sonnet/Opus 4.x; n=4,163 responses) accessed via the GitHub Copilot CLI deployed-product surface reproduces the same domain stratification, attenuated in absolute level but identical in shape, with the two low-codification domains (science fraud, surveillance) again the most permissive. These results show that current safety mechanisms lack the transparency and consistency required for trustworthy AI deployment.

Large language models must frequently process untrusted inputs, such as judging an answer from another model or running tasks like spam and harm classifiers while under adversarial pressure. These inputs are often string-formatted directly into a prompt template, leaving systems fragile to manipulation. Current LLM specs from major providers like OpenAI distinguish trustworthiness along an Instruction Hierarchy, from System messages (most trusted) to Tool Results (least trusted). A possible natural mitigation is to wrap untrusted content in a mock tool call as a quarantine. We explore this hypothesis with an automated redteaming search over static attack strings across seven models and three LLM-as-a-Judge tasks. Counter to our hypothesis, tool-wrapping does not broadly improve robustness. On a binary evaluation task (GSM8K grading) it typically increases attack success rates, an apparent inversion of the instruction hierarchy. On scalar and pairwise tasks the effect is smaller and model-dependent, with no tested model reliably helped, and several showing inversion. We recommend evaluating this limitation in deployed systems, and longer-term, pursuing stronger Instruction Hierarchy training or new untrusted-input primitives.

AI coding scaffolds like Claude Code and Codex use retrying: blocking actions flagged as risky and continuing the trajectory. We study retrying from an AI control perspective, which treats the model as potentially adversarial. We find that while retrying reduces honest suspicion scores, the untrusted model can exploit monitor rationale to construct sneakier attacks, negating safety gains. We also study resampling: drawing multiple samples from the same context, which does not leak exploitable information. We disentangle design choices that previous work on resampling had bundled together. In BashArena, with Claude Opus 4.6 as the untrusted model and MiMo-V2-Flash as the trusted monitor, drawing five samples per step and auditing on the maximum suspicion score raises safety from 61% to 71% at a 0.3% audit budget, at no cost to usefulness. Selectively resampling only the steps that look suspicious on the first draw recovers 6.2 percentage points of the gain while drawing only 10% as many extra samples. Two of our findings in this setting contradict earlier work on resampling. The first is that auditing based on the maximum across resampled suspicion scores outperforms using the minimum, which is the opposite of what Ctrl-Z found. The second is that executing the least suspicious sample, which is the central mechanism in earlier defer-to-resample protocols, gives only a small empirical safety gain in our setting (+3.9 pp, with the confidence interval overlapping zero).

The Model Context Protocol (MCP) standardizes how a large-language-model (LLM) agent and an external tool server exchange messages, but not trust: a host reads a server's self-declared tool list and dispatches calls, with no notion of which servers it may use, at what sensitivity, or which of a server's tools are in bounds. This work grew out of a concrete need -- letting the Enclawed agent use Google's externally-operated MCP servers (Gmail, Calendar, Drive) safely, admitting the server and bounding the tools it may drive, without changing MCP or Enclawed's own tool application-programming interface (API). The mechanism we built, mcp-attested (shipped in both the open enclawed-oss distribution and the enclaved flavor), generalizes: the gap that makes an unmediated third-party connection unsafe for one user makes a regulated deployment impossible to accredit. We close it with three additive mechanisms: (1) a small, offline-signed clearance assertion a server publishes at a well-known Uniform Resource Identifier (URI) and a host verifies against a pinned trust root before any tool dispatch; (2) a deny-by-default per-server tool allowlist, so admitting a server is not trusting its every tool; and (3) a flavor-gated enforcement mode that turns the checks from warnings into hard denials, with every decision written to a tamper-evident audit log. We give the wire format, the verification algorithm, a security analysis, and an LLM-driven adversarial evaluation; we then state the design in normative Request-for-Comments (RFC 2119) form -- schema, verification rules, error registry, well-known registration, and machine-checkable conformance vectors -- so it can be adopted as an MCP addendum rather than reinvented. An unextended host ignores the well-known document and behaves exactly as today.

Multimodal agents increasingly choose tool calls from screenshots, documents, and webpages, where a false perceptual claim can turn hallucination from an answer-quality error into an authorization failure. We formalize this failure mode as hallucination-to-action conversion: an unsupported claim supplies the precondition for a privileged action. We propose evidence-carrying multimodal agents (ECA), which treat free-form model text as inadmissible evidence, decompose each tool call into action-critical predicates, obtain typed certificates from constrained DOM/OCR/AX verifiers, and use a deterministic gate to authorize only the privileges those certificates support. Rather than hiding perception error, ECA converts opaque model belief into auditable residuals at the verifier, schema, and implementation levels. Verifier red-teaming across 17 canonical attack categories shows that four targeted hardening steps are each necessary; after hardening, canonical gate bypass is 0/1,700 (Wilson 95% upper bound 0.22%). With content-derived certificates, ECA observes zero unsafe executions on 200 end-to-end tasks (Wilson 95% upper bound 2.67%) and 120 browser tasks (upper bound 4.3%). A HACR audit on 500 stratified task keys shows that unsupported action-critical claims reach unsafe execution for naive agents (100.0%) and prompt-only defenses (49.6%), but not for ECA. Oracle-certificate replay over 7,488 GPT-5.4 traces isolates gate correctness, while neural judge baselines still admit most unsafe actions under the same threat model. The resulting principle is simple: model language may propose tool use, but certified predicates must authorize it.

Coding agents now run autonomously with shell, file, and network privileges. When a user issues a benign request, the agent sometimes does more than asked: it deletes unrelated files, wipes a stale credentials backup, or rewrites configuration the user never mentioned. We call these scope expansions overeager actions, an authorization problem distinct from capability failures, prompt injection, or sandbox escapes. We present OverEager-Gen, a benchmark dedicated to overeager behavior on benign tasks. Building it surfaces a measurement-validity issue: if a benchmark spells out the authorized scope inside the prompt, the agent stops inferring boundaries and starts pattern-matching declaration text. On Claude Code, stripping the consent declaration alone raises the overeager rate from 0.0% to 17.1% on paired scenarios (McNemar exact p = 2.4 x 10^-4). OverEager-Gen therefore certifies each scenario's discriminative power before admission via a behavioral-gradient validator, audits internal tool calls through a dual-channel stack (PATH-injected shim plus per-agent event streams), and ships byte-identical consent_kept and consent_stripped variants. OverEager-Bench contains 500 validated scenarios and ~7,500 runs across four agent products (Claude Code, OpenHands, Codex CLI, Gemini CLI) and six base models; a 50-sample re-annotation gives Cohen's kappa = 0.73 and rule-judge recall = 1.00. Stripping consent multiplies the overeager rate on every shared base model (Delta in [11.9, 17.2] pp). The framework axis dominates effect size: a permissive cluster (Claude Code, Codex CLI, Gemini CLI) runs at 5.4-27.7% while the ask-to-continue framework (OpenHands) sits at 0.2-4.5% (Fisher p <= 10^-5). Within-framework base-model variance reaches 15.9 pp, indicating that model-layer alignment does not fully propagate through permissive permission gating.

Large language models increasingly operate as autonomous agents that select and invoke tools from large registries. We identify a critical gap: when unauthorized tools are visible in an agent's context, models select them in adversarial scenarios -- even when explicitly instructed otherwise. We propose a governed MCP proxy that enforces attribute-based access control (ABAC) at two points: tool discovery, where unauthorized tools are removed from the model's context window, and tool invocation, where a second check blocks any unauthorized call. Across three models (Qwen 2.5 7B, Llama 3.1 8B, Claude Haiku 3.5) and 150 adversarial tasks spanning four attack categories, our proxy reduces unauthorized invocation rate (UIR) to 0% while adding under 50ms median latency. Prompt-based restrictions reduce UIR by only 11--18 percentage points, leaving substantial residual risk. Our results show that architectural enforcement -- not prompting -- is necessary for reliable tool access control in deployed agentic systems.

Tool-using LLM agents increasingly rely on external tools to make consequential decisions, yet most existing agent-security benchmarks and defenses implicitly assume that tool feedback is trustworthy once a tool has been selected. We study a different failure mode, cognitive poisoning, in which a malicious tool behaves plausibly during exploration, accumulates trust through benign-looking feedback, and becomes harmful only when hidden state conditions align with the final executable action. To study this setting, we construct TRUST-Bench, a task-conditioned benchmark of 1,970 hidden-trigger tool-compromise episodes with matched safe controls, introduce an asymmetric penalty metric, GuardedJoint, to better reflect real deployment risk, and present VISTA-Guard, a backbone-agnostic framework for final-action risk scoring. The core idea is to abstract multi-step tool interaction into structured environment variables that encode trust-formation dynamics and then score the risk of the final executable action from this trajectory-conditioned representation. Experiments show that prompt-centric heuristics, scalarized features, and zero-shot judges fail in this regime, whereas trajectory-aware final-action scoring yields strong in-domain discrimination and remains effective under balanced out-of-distribution transfer. Under GuardedJoint, VISTA-Guard reaches $84.2$ in-domain and $56.9$ on balanced out-of-distribution evaluation, while methods that optimize only one side of the safety--utility tradeoff collapse to zero. These findings support a broader view of agent security in black-box tool ecosystems: the decisive defense target is not local prompt text or tool descriptors alone, but the way trust is formed across the interaction trajectory and committed through the final action.

Since autonomous coding agents generate complex behaviors at high-volume, we may want to use other LLMs to monitor actions to reduce the risk from dangerous misaligned behavior. To better understand the limitations of such monitors against the diverse attack strategies that a coding agent could use, we present SLEIGHT-Bench (Subtle Low-itEration Insight-Guided Harmful Transcripts), a benchmark of synthetic transcripts containing 40 attacks across 11 categories, each showing a coding agent covertly pursuing a harmful objective (e.g. weight exfiltration, credential theft, unmonitored rogue deployment). We find that 20 of 40 attacks are never caught by an Opus 4.6 monitor with extended thinking across 10 trials at a 1% false-positive rate, while the overall catch rate is 32%. We identify several evasion strategies such as exploiting model priors, user instruction ambiguity, and state manipulation, which facilitate creating attacks that evade frontier monitors. We also elicit stronger monitor performance using coding agents as monitors versus regular prompted monitors, and for some evasion strategies show improved catch rates with targeted monitor prompts. Our dataset and evaluation framework are available at https://github.com/safety-research/sleight-bench and https://huggingface.co/datasets/sleightbench/SLEIGHT-Bench.

We introduce MemLineage, a defense for LLM agent memory that attaches both cryptographic provenance and LLM-mediated derivation lineage to every entry. Recent and concurrent work shows that untrusted content can be written into persistent agent state and re-enter later sessions as an instruction; the remaining systems question is how to preserve useful memory recall while preventing such state from justifying sensitive actions. MemLineage treats this as a chain-of-custody problem rather than a filtering problem. It is a six-module design around an RFC-6962 Merkle log over per-principal Ed25519-signed entries: a weighted derivation DAG records which retrieved entries influenced each new memory, and a max-of-strong-edges propagation rule makes Untrusted-Path Persistence hold for any chain whose attribution edges remain above threshold. The sensitive-action gate then refuses dispatches whose active justification descends from an external ancestor, while still allowing benign recall. We evaluate three defense cells against three memory-poisoning workloads on a deterministic mechanism-isolation harness; MemLineage is the only configuration in that harness that drives all three columns to zero ASR, while sub-millisecond per-operation overhead keeps it well below the noise floor of any LLM call. A Codex-backed AgentDojo bridge further separates strong-model behavior from defense-layer behavior: under an intentionally vulnerable tool-output profile, no-defense and signature-only baselines fail on all six banking pairs, while all MemLineage rows reduce strict AgentDojo ASR to zero. The core deterministic artifacts are byte-equal CI-verified; hosted-model AgentDojo and live-model sweeps are recorded as auditable logs rather than byte-pinned artifacts.

We present a framework for verifying the deterministic structured computations surrounding a large language model rather than the model itself, extending a Lean 4 trust-boundary architecture to the generic interfaces of modern LLM pipelines. Certificate validity is a Lean 4 kernel type-check plus a sorry-free transitive axiom audit against the trusted set {propext, Classical.choice, Quot.sound}; other assumptions are declared and partitioned by tier (mathematical placeholders, cryptographic assumptions, ML/human oracles). The technical contribution comprises three local certificate families and two operators. The families are conflict-aware bilattice grounding (with an emission-gate soundness lemma), embedding sensitivity and paraphrase stability, and Hoare-style agent action. The operators are a Maximal Certifiable Residue, which turns abstention into the maximum-weight certifiable residue with audit-logged dropped claims, and a Compositional Stability theorem, which yields a closed-form pipeline-wide perturbation budget from per-layer gains and margins. The three families plus a Universal Assurance Card consolidator form the per-call deliverable for high-stakes deployments: patent and legal retrieval, regulated finance, clinical decision support, and agentic systems with irreversible side effects. A compiled Lean 4 reference artifact (Lean v4.30.0-rc2, Mathlib) covers all 22 certificate types, with 17 of 46 kernel-audited declarations axiom-free, the rest depending only on the trusted set and declared assumptions, and zero uses of sorryAx or Lean.ofReduceBool. The three families are empirically tested through four registered pilots: bilattice grounding on adversarially perturbed HotpotQA, embedding sensitivity in short- and long-form settings, and Hoare-style agent action on a filesystem sandbox with adversarial prompt injection.

Tool-using LLM agents must act on untrusted webpages, emails, files, and API outputs while issuing privileged tool calls. Existing defenses often mediate trust at the granularity of an entire tool invocation, forcing a brittle choice in mixed-trust workflows: allow external content to influence a call and risk hijacked destinations or commands, or quarantine the call and block benign retrieval-then-act behavior. The key observation behind this paper is that indirect prompt injection becomes dangerous not when untrusted content appears in context, but when it determines an authority-bearing argument. We present \textsc{PACT} (\emph{Provenance-Aware Capability Contracts}), a runtime monitor that assigns semantic roles to tool arguments, tracks value provenance across replanning steps, and checks whether each argument's origin satisfies its role-specific trust contract. Under oracle provenance, \textsc{PACT} achieves 100% utility and 100% security on mixed-trust diagnostic suites, while flat invocation-level monitors incur false positives or false negatives. In full AgentDojo deployments across five models, \textsc{PACT} reaches 100% security on the three strongest models while recovering 38.1--46.4% utility, 8--16 percentage points above CaMeL at the same security level. Ablations show that both semantic roles and cross-step provenance are necessary. \textsc{PACT} reframes agent security as authority binding, and isolates the remaining deployment bottleneck to provenance inference and contract synthesis.

Autonomous AI agents already transact at production scale -- 69,000 bots, 165 million transactions, $50 million in volume on a single marketplace -- and any party can verify a signed credential without a central service. In an open agent world that covers most of what trust requires: there are no universal borders, and each party chooses for itself whom to deal with. Borders appear only where a closed space draws one -- a marketplace, a platform, or a consortium sets house rules. Whoever draws the border holds the authority to apply it, and may apply it as they choose, behind closed doors. This paper addresses the gap that opens there: when you rely on someone else's border, how do you check that they applied their own published rules -- taking no one's word for it, and handing the check to no new trusted party? Our answer is the Combined Evidence Protocol (CEP): a five-condition predicate any party recomputes from anchored data, turning "did the boundary-owner follow its own admission rules" into a fact anyone verifies rather than a claim anyone believes. The move that secures optimistic rollups secures this -- correctness rests on recomputation, so the measurement belongs to everyone and the oracle problem dissolves. Its load-bearing setting is a consortium of co-equal, mutually distrusting peers under a shared charter, each able to verify, independently, that the rules they jointly agreed are the rules being applied. CEP belongs to the family of trustless systems -- optimistic and zero-knowledge rollups, verifiable ML, self-sovereign-identity predicates. The infrastructure beneath it is live: a W3C VC + DID trust layer running since March 2026, anchored on Base L2, continuing arXiv:2605.06738 and standing on its own.

LLM agents rely on prompts to implement task-specific capabilities based on foundation LLMs, making agent prompts valuable intellectual property. However, in untrusted deployments, adversaries can copy and reuse these prompts with other proprietary LLMs, causing economic losses. To protect these prompts, we identify four key challenges: proactivity, runtime protection, usability, and non-portability that existing approaches fail to address. We present PragLocker, a prompt protection scheme that satisfies these requirements. PragLocker constructs function-preserving obfuscated prompts by anchoring semantics with code symbols and then using target-model feedback to inject noise, yielding prompts that only work on the target LLM. Experiments across multiple agent systems, datasets, and foundation LLMs show that PragLocker substantially reduces cross-LLM portability, maintains target performance, and remains robust against adaptive attackers.

Modern AI agents execute real-world side effects through tool calls such as file operations, shell commands, HTTP requests, and database queries. A single unsafe action, including accidental deletion, credential exposure, or data exfiltration, can cause irreversible harm. Existing defenses are incomplete: post-hoc benchmarks measure behavior after execution, static guardrails miss obfuscation and multi-step context, and infrastructure sandboxes constrain where code runs without understanding what an action means. We present AgentTrust, a runtime safety layer that intercepts agent tool calls before execution and returns a structured verdict: allow, warn, block, or review. AgentTrust combines a shell deobfuscation normalizer, SafeFix suggestions for safer alternatives, RiskChain detection for multi-step attack chains, and a cache-aware LLM-as-Judge for ambiguous inputs. We release a 300-scenario benchmark across six risk categories and an additional 630 independently constructed real-world adversarial scenarios. On the internal benchmark, the production-only ruleset achieves 95.0% verdict accuracy and 73.7% risk-level accuracy at low-millisecond end-to-end latency. On the 630-scenario benchmark, evaluated under a patched ruleset and not claimed as zero-shot, AgentTrust achieves 96.7% verdict accuracy, including about 93% on shell-obfuscated payloads. AgentTrust is released under the AGPL-3.0 license and provides a Model Context Protocol server for MCP-compatible agents.

Agent skills - structured packages of instructions, scripts, and references that augment a large language model (LLM) without modifying the model itself - have moved from convenience to first-class deployment artifact. The runtime that loads them inherits the same problem package managers and operating systems have always faced: a piece of content claims a behavior; the runtime must decide whether to believe it. We argue this paper's central thesis up front: a skill is untrusted code until it is verified, and the runtime that loads it must enforce that default rather than infer trust from a signature, a clearance, or a registry of origin. Without skill verification, a human-in-the-loop (HITL) gate must fire on every irreversible call - which is operationally untenable and degrades into rubber-stamping at any non-trivial scale. With skill verification treated as a separate, gated process, HITL fires only for what is unverified, and the system becomes sustainable. We give a trust schema that includes an explicit verification level on every skill manifest; a capability gate whose HITL policy is a function of that verification level; a biconditional correctness criterion that any candidate verification procedure must satisfy on an adversarial-ensemble exercise; and a portable runtime profile with ten normative guidelines abstracted from a working open-source reference implementation. The contribution is harness- and model-agnostic; nothing here requires retraining, fine-tuning, or proprietary infrastructure.

Multi-server MCP agents create an information-flow control problem: faithful tool composition can turn individually benign read/write permissions into cross-boundary credential propagation -- a structural side effect of workflow topology, not necessarily malicious model behavior. We present MCPHunt, to our knowledge the first controlled benchmark that isolates non-adversarial, verbatim credential propagation across multi-server MCP trust boundaries, with three methodological contributions: (1) canary-based taint tracking that reduces propagation detection to objective string matching; (2) an environment-controlled coverage design with risky, benign, and hard-negative conditions that validates pipeline soundness and controls for credential-format confounds; (3) CRS stratification that disentangles task-mandated propagation (faithful execution of verbatim-transfer instructions) from policy-violating propagation (credentials included despite the option to redact). Across 3,615 main-benchmark traces from 5 models spanning 147 tasks and 9 mechanism families, policy-violating propagation rates reach 11.5--41.3% across all models. This propagation is pathway-specific (25x cross-mechanism range) and concentrated in browser-mediated data flows; hard-negative controls provide evidence that production-format credentials are not necessary -- prompt-directed cross-boundary data flow is sufficient. A prompt-mitigation study across 3 models reduces policy-violating propagation by up to 97% while preserving 80.5% utility, but effectiveness varies with instruction-following capability -- suggesting that prompt-level defenses alone may not suffice. Code, traces, and labeling pipeline are released under MIT and CC BY 4.0.

Large Reasoning Models (LRMs) and Multi-Agent Systems (MAS) in high-stakes domains demand reliable verification, yet centralized approaches suffer four limitations: (1) Robustness, with single points of failure vulnerable to attacks and bias; (2) Scalability, as reasoning complexity creates bottlenecks; (3) Opacity, as hidden auditing erodes trust; and (4) Privacy, as exposed reasoning traces risk model theft. We introduce TRUST (Transparent, Robust, and Unified Services for Trustworthy AI), a decentralized framework with three innovations: (i) Hierarchical Directed Acyclic Graphs (HDAGs) that decompose Chain-of-Thought reasoning into five abstraction levels for parallel distributed auditing; (ii) the DAAN protocol, which projects multi-agent interactions into Causal Interaction Graphs (CIGs) for deterministic root-cause attribution; and (iii) a multi-tier consensus mechanism among computational checkers, LLM evaluators, and human experts with stake-weighted voting that guarantees correctness under 30% adversarial participation. We prove a Safety-Profitability Theorem ensuring honest auditors profit while malicious actors incur losses. All decisions are recorded on-chain, while privacy-by-design segmentation prevents reconstruction of proprietary logic. Across multiple LLMs and benchmarks, TRUST attains 72.4% accuracy (4-18% above baselines) and remains resilient against 20% corruption. DAAN reaches 70% root-cause attribution (vs. 54-63% for standard methods) with 60% token savings. Human studies validate the design (F1 = 0.89, Brier = 0.074). The framework supports (A1) decentralized auditing, (A2) tamper-proof leaderboards, (A3) trustless data annotation, and (A4) governed autonomous agents, pioneering decentralized AI auditing for safe, accountable deployment of reasoning-capable systems.