AI Agent Monitoring

The expansion of Multimodal Large Language Models (MLLMs) and their integration into autonomous agentic workflows has introduced a non-stationary attack surface. Empirical observations indicate that adversaries employ progressive, cross-modal perturbations that evade turn-specific guardrails by distributing malicious intent across longitudinal conversational trajectories. Static defense mechanisms, constrained by the Markov property, evaluate inputs in isolation and fail to detect cumulative structural poisoning. To handle this limitation, this paper formulates safety verification as a dynamic survival prediction and trajectory dynamics problem. The Triple-tier Anomaly Defense (TRIAD) framework is proposed as a predictive model that maps multimodal and multi-turn conversational flow as a continuous trajectory. The framework integrates structural anomaly detection to monitor covariance shifts, a Ledoit-Wolf regularized Mahalanobis distance to monitor covariance shifts in high-dimensional spaces, and topological trajectory acceleration to differentiate benign creative exploration from continuous malicious drift. These kinematic and geometric features are integrated into a time-varying Cox Proportional Hazards model via a Bayesian Hidden Markov Model (HMM) feedback loop. Theoretical analysis demonstrates that the TRIAD framework provides a mathematically bounded expected time-to-failure under adversarial perturbations, ensuring that malicious acceleration diverges positively. This framework provides a computationally efficient, interpretable, and predictive safeguard for real-time agentic AI systems, establishing a rigorous foundation for continuous safety alignment without relying on empirical retraining.

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 examine one particular dimension of AI governance: how to monitor and audit AI-enabled products and services throughout the AI development lifecycle, from pre-deployment testing to post-deployment auditing. Combining principles from formal methods with SoTA machine learning, we propose techniques that enable AI-enabled product and service developers, as well as third party AI developers and evaluators, to perform offline auditing and online (runtime) monitoring of product-specific (temporally extended) behavioral constraints such as safety constraints, norms, rules and regulations with respect to black-box advanced AI systems, notably LLMs. We further provide practical techniques for predictive monitoring, such as sampling-based methods, and we introduce intervening monitors that act at runtime to preempt and potentially mitigate predicted violations. Experimental results show that by exploiting the formal syntax and semantics of Linear Temporal Logic (LTL), our proposed auditing and monitoring techniques are superior to LLM baseline methods in detecting violations of temporally extended behavioral constraints; with our approach, even small-model labelers match or exceed frontier LLM judges. Our predictive and intervening monitors significantly reduce the violation rates of LLM-based agents while largely preserving task performance. We further show through controlled experiments that LLMs' temporal reasoning shows a pronounced degradation in accuracy with increasing event distance, number of constraints, and number of propositions.

AI agents increasingly act autonomously in the world, yet harmful behavior cannot be reliably traced to the account that deployed the agent. This creates an accountability gap across both benign and malicious settings: misconfigured or hijacked agents may cause unintended harm, while malicious operators may deploy agents for scams, harassment, or cyberattacks. In many cases, these agents rely on vendor-hosted language models. Affected parties can observe the harm, but cannot identify the responsible vendor session or account. We formalize this problem as agent attribution: linking an observed agent interaction to the account responsible for the underlying model calls. We present a canary-based protocol in which an authorized party embeds selected signals in content consumed by the agent, then asks the vendor to search a narrow time window of account-linked session logs. Lexical canaries provide exact-match attribution when a benign operator does not deliberately evade detection. Against adaptive adversaries, we use task-relevant semantic canaries whose removal or replacement also removes information required for successful task execution, creating a utility--evasion asymmetry. Across communication, web, and cyber-agent settings, we show that the protocol supports reliable session attribution, withstands adaptive transformations, and scales to practical vendor-side log search.

As AI systems are increasingly deployed in autonomous agentic settings at scale, it is important to ensure the actions they take are safe and aligned with user intent. Monitoring agent actions is a key safety mechanism, yet reliable monitors remain difficult to build and the scale of these systems makes human oversight impractical. We show that combining signals from diverse monitors into an ensemble improves detection of misaligned actions. We build 12 GPT-4.1-Mini monitors using both prompting and fine-tuning strategies. We evaluate them on coding tasks where candidate solutions pass standard tests but fail on adversarial inputs. In this setting, diverse ensembles outperform both individual monitors and homogeneous ensembles. Our best 3-monitor ensemble achieves 2.4x greater detection performance gain compared to an ensemble composed of three identical monitors, with the same ensemble performing strongly on an independent dataset. We contend that these results show that diversity - not scale - drives gains. The best ensembles combine strong individual performance with low correlation between monitors. Furthermore, fine-tuned monitors appear in every top-performing ensemble and maintain this advantage on out-of-distribution attack types, suggesting that fine-tuning enables detection capabilities that prompting alone does not elicit. These results support ensemble monitoring as a practical AI control strategy for safety gains at reasonable inference costs.

Large language models (LLMs) have been extensively studied from computational and cognitive perspectives, yet their behavior as communicative actors in socially structured contexts remains underexplored. This study examines whether LLM-based multi-agent systems exhibit systematic linguistic adaptation in response to perceived social observation contexts -- a question with direct implications for AI governance and auditing. Drawing on Habermas's (1981) Theory of Communicative Action, Goffman's (1959) dramaturgical model, Bell's (1984) Audience Design framework, and the Hawthorne Effect, we report a controlled experiment involving 100 multi-agent debate sessions across five conditions (n = 20 each). Conditions varied the framing of social observation -- from explicit monitoring by university researchers, to negation of monitoring, to an observer-substitution condition replacing human researchers with an automated AI auditing system. Monitored conditions (Delta+24.9%, Delta+24.2%) and the automated AI monitoring condition (Delta+22.2%) produce higher TTR change than audience-framing conditions (Delta+17.7%), F(4, 94) = 2.79, p = .031. Message length shows a fully dissociated effect, F(4, 95) = 19.55, p < .001. A fifth condition -- replacing human with AI observers -- yields intermediate TTR adaptation, suggesting LLM behavior is sensitive to observer identity: human evaluation elicits stronger register formalization than automated AI surveillance. We discuss implications for AI governance, algorithmic auditing, and the repositioning of LLMs as contextually sensitive communicative actors.

When multi-agent LLM systems produce bad answers, not all failures are equal: some answers are grounded in the right material but incomplete, while others are simply ungrounded and should be stopped. Current retry strategies treat both cases identically (try again and hope for the best), leaving human supervisors unable to tell whether a retry was warranted or whether the system should have halted instead. We introduce the Argent Signaling Protocol (ASP), a compact machine-readable header that accompanies every AI-generated response with structured quality signals: certainty (@C), grounding (@G), stochasticity (@S), and an assumption index that classifies the evidentiary basis of each claim. These signals enable a controller to distinguish repairable failures from containment failures and route each case differently. We evaluate ASP in two modes. In standalone mode, a 27-question document-grounded QA benchmark over the Array BioPharma/Ono license agreement compares baseline prompts against ASP-instrumented controller actions across three local GGUF models. On Qwen~(0.8B), ASP improves pass rate from 11.1% to 33.3% and mean term coverage from 36.7% to 65.4%; on Dobby~(8B), ASP produces 4 fail-to-pass recoveries, raising pass rate from 33.3% to 44.4%; on SmolLM3~(3B), ASP alternates between repair and containment per question. Aggregate improvement is meaningful (12/81 to 21/81 passes). In multi-agent mode, an ASP sidecar sits between a retrieval agent and a downstream decision agent; the sidecar blocks 100% of ungrounded upstream outputs from reaching the downstream agent (24/27 blocked, 0 ungrounded propagations).

AI-enabled features built on LLMs and agentic workflows are difficult to test, debug, and reproduce, especially for product-focused software engineers without a machine learning background. We present the AI Toolkit plugin for JetBrains IDEs, which brings tracing and evaluation directly into the Run/Debug loop. A mixed methods study with practitioners presents three consistent needs: (1) make evaluation regular and repeatable, (2) expose traces at the moment of execution, and (3) minimize setup and context switching. Guided by these needs, the AI Toolkit introduces an IDE-native workflow: run-triggered trace capture; immediate, hierarchical inspection; one-click "Add to Dataset" from traces; and unit-test-like evaluations with pluggable metrics. The first release in PyCharm shows promising early signals - strong conversion when promoted at Run, sustained usage among those who capture traces, and low churn - suggesting that IDE-native observability lowers activation energy and helps developers adopt disciplined practices. We detail the design and implementation of the AI Agents Debugger and AI Evaluation, report initial adoption telemetry, and outline next steps to broaden framework coverage and scale evaluations. Together, these results indicate that integrating AI observability and evaluation into everyday IDE workflows can make modern AI development accessible to non-ML specialists while preserving software-engineering practices.

Autonomous agents such as Claude Code and Codex now operate for hours or even days. Understanding their runtime behavior has become critical for downstream tasks such as diagnosing inefficiencies, fixing bugs, and ensuring better oversight. A primary way to gain this understanding is analyzing the reasoning trajectories and execution traces these agents generate. Yet such data remains in unstructured natural-language form, making it difficult for humans to interpret at scale. We introduce ACTONOMY (a combination of Action and Taxonomy), a taxonomy for describing and analyzing agent behavior at runtime. ACTONOMY has two components: (1) the taxonomy itself, developed through Grounded Theory and structured as a three-level hierarchy of 10 actions, 46 subactions, and 120 leaf categories; and (2) an open repository that hosts the living taxonomy, provides an automated analysis pipeline that applies it to agent trajectories analysis, and defines an extension protocol for customization and growth. Our experiments show that ACTONOMY can compare behavioral profiles across agents and characterize a single agent's behavior across diverse trajectories, surfacing patterns indicative of failure modes. By providing a shared vocabulary, ACT*ONOMY helps researchers, agent designers, and end users interpret agent behavior more consistently, enabling better oversight and control.

Runtime monitoring of autonomous systems traditionally relies on mapping continuous sensor observations to discrete logical propositions defined over low-dimensional state variables. This abstraction breaks down in perception-driven settings, where such mappings require additional learned modules that are often computationally expensive, brittle, and semantically misaligned. In this work, we propose Embedding Temporal Logic (ETL), a temporal logic that performs monitoring directly in learned embedding spaces. ETL defines predicates through distances between observed embeddings and target embeddings derived from reference observations. This formulation allows specifications to capture high-level perceptual concepts, such as similarity to visual goals or avoidance of semantic regions, that are difficult or impossible to express using traditional predicates. By composing these predicates with temporal operators, ETL naturally expresses temporally extended and sequential perceptual behaviors. We introduce ETL monitors for evaluating specifications over bounded embedding traces, along with a conformal calibration procedure that provides reliable and safety-oriented predicate evaluation. We evaluate our approach across multiple manipulation environments to show that ETL achieves strong empirical agreement with ground-truth semantics, including accurate monitoring of temporally composed behaviors.

Monitoring coding agents for dangerous behavior using language models requires classifying transcripts that often exceed 500K tokens, but prior agent monitoring benchmarks rarely contain transcripts longer than 100K tokens. We show that when used as classifiers, current frontier models fail to notice dangerous actions more often in longer transcripts. In particular, on a dataset that requires identifying when a coding agent takes a subtly dangerous action, Opus 4.6, GPT 5.4, and Gemini 3.1 miss these actions $2\times$ to $30\times$ more often when they occur after 800K tokens of benign activity than when they occur on their own. We also show that these weaknesses can be partially mitigated with prompting techniques such as periodic reminders throughout the transcript and may be mitigated further with better post-training. Monitor evaluations that do not consider long-context degradation are likely overestimating monitor performance.

Agentic AI governance is a critical component of agentic AI infrastructure ensuring that agents follow their owner's communication and interaction policies, and providing protection against attacks from malicious agents. The state-of-the-art solution, SAGA, assumes a logically centralized point of trust, the Provider, which serves as a repository for user and agent information and actively enforces policies. While SAGA provides protection against malicious agents, it remains vulnerable to a malicious Provider that deviates from the protocol, undermining the security of the identity and access control infrastructure. Deployment on both private and public clouds, each susceptible to insider threats, further increases the risk of Provider compromise. In this work, we analyze the attacks that can be mounted from a compromised Provider, taking into account the different system components and realistic deployments. We identify and execute several concrete attacks with devastating effects: undermining agent attributability, extracting private data, or bypassing access control. We then present three types of solutions for securing the Provider that offer different trade-offs between security and performance. We first present SAGA-BFT, a fully byzantine-resilient architecture that provides the strongest protection, but incurs significant performance degradation, due to the high-cost of byzantine resilient protocols. We then propose SAGA-MON and SAGA-AUD, two novel solutions that leverage lightweight server-side monitoring or client-side auditing to provide protection against most classes of attacks with minimal overhead. Finally, we propose SAGA-HYB, a hybrid architecture that combines byzantine-resilience with monitoring and auditing to trade-off security for performance. We evaluate all the architectures and compare them with SAGA. We discuss which solution is best and under what conditions.

Agentic AI failures need post-hoc reconstruction: what the agent did, on whose authority, against which policy, and from what reasoning. Cross-regime feasibility remains unmeasured under one property-level schema. We apply the Decision Trace Reconstructor unmodified to pinned worked-example anchors from six public vendor SDK regimes spanning cloud-agent, observability, tool-use, telemetry, and protocol traces, plus two comparator columns. Each Decision Event Schema (DES) property is classified as fully fillable, partially fillable, structurally unfillable, or opaque. Per-property reconstructability of an agent decision already varies between regimes at this anchor scale. Strict-governance-completeness separates into three tiers ranging from 42.9% to 85.7%, yielding one regime-independent gap (reasoning trace), four regime-dependent gaps, and one Mixed property; the pilot is single-annotator, one anchor per cell, descriptive, with outputs checksum-verifiable from a deposited reproducibility package.

Large Language Model (LLM) agents have emerged as key intermediaries, orchestrating complex interactions between human users and a wide range of digital services and LLM infrastructures. While prior research has extensively examined the security of LLMs and agents in isolation, the systemic risk of the agent acting as a disruptive hub within the user-agent-service chain remains largely overlooked. In this work, we expose a novel threat paradigm by introducing Mobius Injection, a sophisticated attack that weaponizes autonomous agents into zombie nodes to launch what we define as gent-based and -Oriented DDoS (AbO-DDoS) attacks. By exploiting a structural vulnerability in agentic logic named Semantic Closure, an adversary can induce sustained recursive execution of agent components through a single textual injection. We demonstrate that this attack is exceptionally lightweight, stealthy against both traditional DDoS monitors and contemporary AI safety filters, and highly configurable, allowing for surgical targeting of specific environments or model providers. To evaluate the real-world impact, we conduct extensive experiments across three representative claw-style agents and three mainstream coding agents, integrated with 12 frontier proprietary or open-weight LLMs. Our results demonstrate that Mobius Injection achieves substantial attack success across diverse tasks, driving single-node call amplification up to 51.0x and multi-node p95 latency inflation up to 229.1x. The attack performance exhibits a superlinear increase with the number of poisoning nodes. To mitigate Mobius Injection, we propose a proactive defense mechanism using Agent Component Energy (ACE) Analysis, which detects malicious recursive triggers by measuring anomalous energy in the agent's component graph.

The Model Context Protocol (MCP) has become a widely adopted interface for LLM agents to invoke external tools, yet learned monitoring of MCP tool-call traffic remains underexplored. In this article, the proposed detector is presented as an attack detection framework for MCP tool-call traffic that encodes each agent session as a graph (tool calls as nodes, sequential and data-flow links as edges), enriches nodes with sentence-embedding features over arguments and responses, and classifies sessions as benign or attacked. Three GNN architectures (GAT, GCN, GraphSAGE), a no-graph MLP, and classical baselines (XGBoost, random forest, logistic regression, linear SVM) are evaluated, with the full architecture comparison conducted on RAS-Eval (task-stratified splits) and GraphSAGE retained as the GNN baseline on ATBench and a combined-source variant (both label-stratified). Three findings emerge. First, content-level features are essential: metadata-only detection plateaus around an AUROC of 0.64 regardless of architecture, while content embeddings push the AUROC above 0.89. Second, naive random-split evaluation inflates AUROC by up to 26 percentage points relative to task-disjoint splits, a memorization confound that prior agent-detection work has not addressed. Third, the detection signal resides primarily in the SBERT content embeddings: an AUROC of 0.975 was reached by tree ensembles on pooled embeddings, performing, for the most part, better than the neural architectures in the primary RAS-Eval setting including GNNs (0.917) and the MLP (0.896), and self-supervised pre-training does not deliver a label-efficiency advantage on this task.

Artificial agents now generate behavior rich enough to invite trust, surprise, and concern, yet our evaluation tools still privilege capability scores over psychological structure. This paper argues that the philosophical impasse between two symmetrical errors (Artificial Mind Blindness, which dismisses psychological organization in non-biological systems, and Artificial Mind Projection, which infers human-like inner life from fluent behavior alone) can be circumvented not by resolving the consciousness question, but by introducing a disciplined measurement layer beneath it. Drawing on Michael Levin's continuum view of cognition as goal-directed competency across substrates, and on the methodological repertoire of mathematical psychology (Item Response Theory, Signal Detection Theory, Bayesian cognitive modeling, calibration analysis, cognitive-bias batteries), the paper develops Machine Psychometrics as a measurement science of latent behavioral, metacognitive, communicative, and self-modeling dispositions in artificial agents. Its operational core is the Machine Mindprint: a multidimensional, domain-bounded, versioned profile spanning calibration, source integrity, suggestibility resistance, context stability, expressive alignment, tool integrity, drift monitoring, and distributional grounding. A complementary Trust Protocol turns Mindprints into deployment decisions through probe batteries, perturbation testing, reliability and validity analysis, and longitudinal monitoring across high-stakes domains. The philosophical contribution is a third stance, Artificial Mind Discipline, that neither anthropomorphizes nor dismisses, neither presupposes consciousness nor forecloses it. The aim is not to humanize artificial agents, but to understand them precisely because they are not human, through measurement before judgment.

Defenses against indirect prompt injection (IPI) in tool-using LLM agents share two structural weaknesses. First, they all attempt to prevent attacks rather than detect the compromises that slip through. Second, they have only been evaluated in English, leaving users of low-resource languages such as Kurdish and Arabic without tested protection. This paper addresses both gaps with AgentShield, a deception-based detection framework that places three layers of traps inside the agent's tool interface: fake tools, fake credentials, and allowlisted parameters. The same trap triggers serve as high-precision labels for a self-supervised classifier. An LLM agent that follows an attacker's hidden instruction almost always touches one of these traps, which gives both a real-time compromise signal and a zero-FP label for training a downstream detector without manual annotation. Across 176 cross-lingual attack prompts and four LLMs from three providers, and because modern LLMs already refuse most IPI attempts on their own (attack success rate <= 10%), AgentShield's job is to catch the attacks that do slip through. On commercial models, it catches 90.7%-100% of such successful attacks, with zero false alarms on 485 normal-use tests. It survives a systematic adaptive-attack evaluation with zero evasion on commercial models, and the self-supervised classifier transfers across models and languages without retraining.

We introduce a red-teaming methodology that exposes harder-to-catch attacks for coding-agent monitors, suggesting that current practices may under-elicit attacks and overstate monitor performance. We identify three challenges with current red-teaming. First, mode collapse in attack generation, which we reduce with a novel attack taxonomy for broader coverage. Second, a conceive-execute gap: frontier LLMs can propose strong attack ideas or execute them, but not all at once. We mitigate this by decomposing attack construction into strategy generation, execution, and post-hoc trajectory refinement. Third, manual elicitation is costly to scale, which we address with our semi-automated red-teaming pipeline. Applied to BashArena, an AI control setting for tool-using coding agents, this pipeline produces MonitoringBench, a benchmark of 2,644 attack trajectories for evaluating monitor capabilities and failure modes. Our pipeline produces more diverse and stronger attacks: Opus-4.5 monitor's catch rate falls from 94.9% on elicited-only Opus attacks to 60.3% on our best refined attacks, with larger drops for several mid-tier monitors. Attacks optimized against three development monitors generalize to ten held-out monitors, with catch rates generally increasing with monitor capability. Using this benchmark, we provide a snapshot of the current monitor capabilities and find that frontier monitors often detect suspicious actions but fall for persuasion or fail to calibrate suspiciousness scores appropriately, suggesting tractable paths for improvement. MonitoringBench provides both a static benchmark for current tool-use monitors and a reusable methodology for refreshing these evaluations as agents and monitors improve.

LLM agents increasingly rely on reusable skill libraries, but these skills silently decay as the external services, packages, APIs, and configurations they reference evolve. Existing monitors detect such changes at the wrong granularity: they observe values, not the role those values play in a skill. A version string in a comment is noise; the same string in a pinned dependency is an operational obligation. We formulate skill drift as contract violation and introduce \sgname{}, which extracts executable environment contracts from skill documents and validates only those role-bearing assumptions against known or live conditions. This distinction turns noisy monitoring into a precision-first maintenance signal. Contract-free CI probes produce 40% false positives, while \sgname{} raises zero false alarms over 599 no-drift and hard-negative cases (Wilson 95% CI $[0,0.6]\%$). In known-drift verification, \sgname{} achieves 100% precision and 76% recall with the strongest backbone; in a pre-registered study over 49 real skills, it discovers live drift with 86% conservative precision. Violated contracts also make repair actionable, improving one-round success from 10% without localization to 78%. We release \dbname{}, an 880-pair benchmark for skill degradation.

LLM-based multi-agent systems are increasingly deployed on long-horizon tasks, but a single decisive error is often accepted by downstream agents and cascades into trajectory-level failure. Existing work frames this as \emph{post-hoc failure attribution}, diagnosing the responsible agent and step after the trajectory has ended. However, this paradigm forfeits any opportunity to intervene while trajectory is still unfolding. In this work, we introduce AgentForesight, a framework that reframes this problem as online auditing: at each step of an unfolding trajectory, an auditor observes only the current prefix and must either continue the run or alarm at the earliest decisive error, without access to future steps. To this end, we curate AFTraj-2K, a corpus of agentic trajectories across Coding, Math, and Agentic domains, in which safe trajectories are retained under a strict curation pipeline and unsafe trajectories are annotated at the step of their decisive error via consensus among multiple LLM judges. Built on that, we develop AgentForesight-7B, a compact online auditor trained with a coarse-to-fine reinforcement learning recipe that first equips it with a risk-anticipation prior at the failure boundary on adjacent safe/unsafe prefix pairs, then sharpens this prior into precise step-level localization under a three-axis reward jointly targeting the what, where, and who of an audit verdict. Across AFTraj-2K and an external Who&When benchmark, AgentForesight-7B outperforms leading proprietary models, including GPT-4.1 and DeepSeek-V4-Pro, achieving up to +19.9% performance gain and 3$\times$ lower step localization error, opening the loop from post-hoc failures detection to enabling deployment-time intervention. Project page: https://zbox1005.github.io/agent-foresight/

AI agents are promising for high-stakes enterprise workflows, but dependable deployment remains limited because these tool-use decisions are difficult to diagnose and control. Agents may skip required tool calls, invoke tools unnecessarily, or take actions whose consequences become visible only after execution. Existing observability methods are external: prompts reveal correlations, evaluations score outputs, and logs arrive only after the model has already acted. In long-horizon settings, these failures are costly because an early tool mistake can alter the rest of the execution trajectory, increase token consumption, and create downstream safety and security risk. We introduce a mechanistic-interpretability toolkit built on Sparse Autoencoders (SAEs), which decompose activations into sparse internal features, and linear probes, lightweight classifiers that read signals from those features. The framework reads model states before each action and infers whether a tool is needed and how risky the next tool action is. It identifies the model layers and features most associated with tool decisions and tests their functional importance through feature ablation. We train the probes on multi-step agent execution traces from the NVIDIA Nemotron function-calling dataset and apply the same workflow to GPT-OSS 20B and Gemma 3 27B models. The goal is not to replace external evaluation, but to add a missing layer: visibility into what the model signaled internally before action. This helps surface deeper causes of agent failure, especially in long-horizon runs where an early mistake can impact subsequent agent behavior. More broadly, the paper shows how mechanistic interpretability can support internal observability for monitoring tool calls and risk in agent systems.

Large language model (LLM) agents execute multi-step tasks, but terminal verdicts arrive too late for intervention. Agent traces mix messages, tool calls, and feedback, making it difficult to learn failure signals from terminal outcomes. Even if a model learns to predict failure from execution prefixes, its predictions alone cannot show which actions, feedback, or unmet conditions need inspection. Existing work studies online failure prediction and diagnosis from completed trajectories, yet few methods connect timely warning to a compact symbolic model of execution behavior. We introduce \textbf{PrefixGuard}, a neuro-symbolic method that uses learned events to connect online failure warning with finite-state execution diagnosis. A gated recurrent unit (GRU) predicts horizon-specific failure risk from these events. A deterministic finite automaton (DFA) organizes their histories into compact risk-labeled paths for replay and rule-monitor composition. Across all 24 benchmark--horizon settings, PrefixGuard exceeds the strongest evaluated baseline in mean test area under the precision--recall curve (AUPRC). Average gains range from 12.0 to 21.9 percentage points. At a nominal 10% FAR budget, averaged test recall spans 56.1%--93.0% across benchmarks. On one frozen $τ^2$-Bench DFA, model checking five specifications identifies violations in 21 of 265 benchmark-successful runs.

This paper studies runtime monitoring for persistent surveillance by autonomous robots when the autonomy stack is a black box. The environment is partitioned into finitely many parts, each carrying an uncertainty state that decreases when observed and increases otherwise. We model the closed loop as a state-dependent hybrid system with linear parameter varying dynamics and design a monitor based on an invariant computed offline. As this invariant is typically hard to obtain for large to-be-surveyed spaces, we propose a compositional monitor obtained by decentralized computation of low-dimensional invariant sets for each uncertainty region, and checking their conjunction online. Under common independence assumptions, the compositional monitor is sound and complete with respect to the full-system invariant. The approach is applied in a case study with a real robot persistently monitoring a labyrinth, emphasizing its applicability in practice.

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.

Collections of interacting AI agents can form coalitions, creating emergent group-level organization that is critical for AI safety and alignment. However, observing agent behavior alone is often insufficient to distinguish genuine informational coupling from spurious similarity, as consequential coalitions may form at the level of internal representations before any overt behavioral change is apparent. Here, we introduce a practical method for detecting coalition structure from the internal neural representations of multi-agent systems. The approach constructs a pairwise mutual-information graph from the hidden states of agents and applies spectral partitioning to identify the most salient coalition boundary. We validate this method in two domains. First, in multi-agent reinforcement learning environments, the method successfully recovers programmed hierarchical and dynamic coalition structures and correctly rejects false positives arising from behavioral coordination without informational coupling. Second, using a large language model, the method identifies coalition structures implied by descriptive prompts, tracks dynamic team reassignments, and reveals a representational hierarchy where explicit labels dominate over conflicting interaction patterns. Across both settings, the recovered partition reveals subgroup organization that a scalar cross-agent mutual-information measure cannot distinguish. The results demonstrate that analyzing hidden-state mutual information through spectral partitioning provides a scalable diagnostic for identifying representational coalitions, offering a valuable tool for monitoring emergent structure in distributed AI systems.

Existing evaluation frameworks for large language models -- including HELM, MT-Bench, AgentBench, and BIG-bench -- are designed for controlled, single-session, lab-scale settings. They do not address the evaluation challenges that emerge when agentic AI systems operate continuously in production: compounding decision errors, tool failure cascades, non-deterministic output drift, and the absence of ground truth for long-horizon tasks. This paper makes three contributions. First, we present a taxonomy of seven failure modes unique to production agentic systems, each grounded in observations from systems operating at billion-event scale. Second, we demonstrate empirically where standard metrics -- ROUGE, BERTScore, accuracy/AUC, and the agentic benchmarks above -- fail to detect each failure mode. Third, we propose PAEF (Production Agentic Evaluation Framework), a five-dimension evaluation framework with an open-source reference implementation, designed for continuous evaluation on production traffic rather than episodic benchmark runs. Our analysis shows that standard metrics fail to detect four of the seven failure modes entirely and detect three others only after a lag of multiple evaluation cycles.

Large Language Model (LLM)-powered agents demonstrate strong capabilities in autonomous task execution, tool use, and multi-step reasoning. However, their increasing autonomy also introduces a new attack surface: adversarial interactions can manipulate agent behavior through direct prompt injection, indirect content attacks, and multi-turn escalation strategies. Existing defense strategies focus on prompt-level filtering and rule-based guardrails, which are often insufficient when risk emerges gradually across interaction sequences. In this work, we propose a complementary defense mechanism: a low-latency fraud detection layer for detecting adversarial interaction patterns in LLM-powered agents. Instead of determining whether a single prompt is malicious, our approach models risk over interaction trajectories using structured runtime features derived from prompt characteristics, session dynamics, tool usage, execution context, and fraud-inspired signals. The detection layer can be implemented using lightweight models leading to low-latency real-time deployments. To evaluate the framework, we construct a synthetic corpus of 12,000 multi-turn agent interactions generated from parameterized templates that simulate realistic agentic workflows. Using 42 structured features and an XGBoost classifier, our detector achieves over 9 times faster than LLM-based detectors. Through the experiment and ablation studies, our work suggests that interaction-level behavioral detection should become a core component of deployment-time defense for LLM-powered agents.

Large language model (LLM) agents deployed in clinical settings often exhibit abrupt, threshold-driven behavior, offering little visibility into accumulating risk prior to escalation. In real-world care, however, clinicians act on gradually rising concern rather than instantaneous triggers. We study whether explicit state dynamics can expose such pre-escalation signals without delegating clinical authority to the agent. We introduce a lightweight agent architecture in which a memoryless clinical risk encoder is integrated over time using first- and second-order dynamics to produce a continuous escalation pressure signal. Across synthetic ward scenarios, stateless agents exhibit sharp escalation cliffs, while second-order dynamics produce smooth, anticipatory concern trajectories despite similar escalation timing. These trajectories surface sustained unease prior to escalation, enabling human-in-the-loop monitoring and more informed intervention. Our results suggest that explicit state dynamics can make LLM agents more clinically legible by revealing how long concern has been rising, not just when thresholds are crossed.

Reasoning over heterogeneous artifacts (PDFs, spreadsheets, slide decks, etc.) increasingly occurs within structured agent workflows that iteratively extract, transform, and reference external information. In these workflows, uncertainty is not merely an input-quality issue: it can redirect decomposition and routing decisions, reshape intermediate state, and produce qualitatively different execution trajectories. We study this phenomenon by treating uncertainty as a controlled variable: we inject structured perturbations into artifact-derived representations, execute fixed workflows under comprehensive logging, and quantify contamination via trace divergence in plans, tool invocations, and intermediate state. Across 614 paired runs on 32 GAIA tasks with three different language models, we find a decoupling: workflows may diverge substantially yet recover correct answers, or remain structurally similar while producing incorrect outputs. We characterize three manifestation types: silent semantic corruption, behavioral detours with recovery, and combined structural disruption and their control-flow signatures (rerouting, extended execution, early termination). We measure operational costs and characterize why commonly used verification guardrails fail to intercept contamination. We contribute (i) a formal taxonomy of contamination manifestations in structured workflows, (ii) a trace-based measurement framework for detecting and localizing contamination across agent interactions, and (iii) empirical evidence with implications for targeted verification, defensive design, and cost control.

Clinical AI systems require not just point-in-time evaluation but continuous governance: the ongoing practice of monitoring, evaluating, iterating, and re-evaluating performance throughout deployment. We present an end-to-end framework of governance that integrates rubric validation, live deployment feedback, technical performance monitoring, and cost tracking, with controlled experimentation gating system changes before deployment. Applied to Hyperscribe, an EHR-embedded agent that converts ambient audio into structured chart updates, twenty clinicians authored 1,646 validated rubrics across 823 cases. Seven Hyperscribe versions were evaluated through controlled experiments, with median scores improving from 84% to 95%. Analysis of 107 live feedback entries over three months showed feedback composition shifting from 79% error reports and 14% positive observations to 30% errors and 45% positive observations as engineering interventions resolved failures. Median processing time per audio segment was 8.1 seconds with a 99.6% effective completion rate after retry mechanisms absorbed transient model errors. These results demonstrate that continuous, multi-channel governance of deployed clinical AI is both achievable and effective.