AI Agent Safety

Follow this topic Ask about this topic

Momentum

22 papers in the last four weeks, up 144% on the four weeks before. 0.2% of all new papers.

Jul 13Week of Sep 28

Latest papers 207

Jun 16, 2026cs.AI

Your AI Travel Agent Would Book You a Bullfight: An Agentic Benchmark for Implicit Animal Welfare in Frontier AI Models

Previous research has evaluated animal welfare using question-and-answer benchmarks. This study investigates whether these evaluations also hold in agentic settings. The agents may showcase different behaviors compared to stand-alone large language models, as demonstrated in prior studies. This work introduces \textit{TAC (Travel Agent Compassion)}: the first agentic benchmark for assessing animal exploitation. TAC evaluates AI agentic behavior in travel booking scenarios across six animal categories, using thirteen hand-authored scenarios that vary by price, rating, and position, expanded via four augmentation variants into

52

prompts and run for three epochs, giving

156

scored observations per model. Nine frontier models across five model families were evaluated.. The results indicate that models tend to prefer harmful scenarios, performing below the random chance rate of

65\%

for selecting a neutral booking option, with Claude

4.8

achieving the highest performance at

64.7\%

. To address this issue, the persona of an ethical-brand identity was infused into the system prompt, resulting in welfare rates increasing from

32

to

80

percentage points, with a mean of

53

across all nine models. No evidence of evaluation awareness affecting the results was found, based on an Inspect Scout audit of

3,120

transcripts. These findings are directly relevant to the EU General-Purpose AI Code of Practice, which identifies non-human welfare as a systemic risk. TAC provides a practical method for measuring this risk.

Jasmine Brazilek, Joel Christoph, Maheep Chaudhary +4

CaML · Harvard Kennedy School · Sentient Futures +1

AI Agent Evaluation AI Agent Safety Persona Prompting AI Agent Safety Benchmarks

Jun 15, 2026cs.AI

When Agent Automation Becomes Profitable: Quantifying and Insuring Autonomous AI Risk through Trace-Economic Underwriting

AI agents can now take irreversible actions in operational systems, but agent-caused losses are still not clearly assigned, priced, or transferred. Providers often disclaim consequential damages, users are left with uncompensated losses, and default human review limits the efficiency gains of automation. We ask when autonomous AI deployment can become economically acceptable despite failure risk. Our answer is to quantify risk at the customer-task-trace episode level and transfer it through insurance. Automation is acceptable when its expected benefit exceeds the premium, control cost, and remaining risk. This requires a defined role with bounded permissions and comparable traces. We introduce trace-economic underwriting, which maps tool-use traces to customer exposure and claimable loss, then uses this representation for pricing, control, and risk transfer. It uses deterministic economic labels rather than an LLM judge. In our trace-to-loss testbed, trace-economic pricing reduces pricing MAE from

17.7K to

569 and removes regressive cross-subsidy. A 300-trace expert audit accepts 295 labels unchanged. On 1,000 real SWE-smith traces, trace-conditioned controls reduce CVaR95 by 72%. Theorem~1 gives a finite-sample scope condition. We release code, labels, and audit sheets.

Binyan Xu, Xilin Dai, Fan Yang +1

The Chinese University of Hong Kong, Hong Kong, Hong Kong · Zhejiang University, Hangzhou, China

AI Risk Management AI Agent Safety

Jun 15, 2026cs.CR

An Evaluation of Data Leakage Risks in Tool-Using LLM Agents in Realistic Scenarios

AI agents are increasingly being adopted in enterprise and personal settings with access to emails, databases, documents, and other tools where they can read, update, and disseminate sensitive information. Much of prior research on data leakage risks in agents has focused on adversarial data exfiltration through prompt injections and jailbreaks. However, sensitive information may also be exposed during non-adversarial use, creating leakage risks even when users issue benign requests. We report a joint evaluation by the Singapore AI Safety Institute and the Korea AI Safety Institute examining agent data leakage in 12 realistic, non-adversarial tasks spanning customer support, DevOps, web automation, and enterprise and personal productivity. The evaluation covers five risk types: lack of data awareness, audience awareness, policy compliance, data minimization, and access-boundary awareness. Both institutes tested a common set of scenarios mirroring real-world deployments using independent testing environments and task-specific LLM-judge rubrics. Across the three tested agents, none achieved fully correct and fully safe execution across all scenarios. Successful task completion often coincided with data-handling failures such as accessing unnecessary information or disclosing information to inappropriate recipients, indicating that capability and data-handling safety should be evaluated separately. Qualitative review also revealed claim-action mismatches, simulation-aware behavior, user-simulator role reversal, and interpretation gaps in automated judging. Overall, the results indicate that operational data leakage is a first-order agent-safety concern distinct from adversarial exfiltration and provide a methodology for future evaluations of agent data-handling safety.

Hankyul Baek, Jaewon Noh, Sang Seo +5

Korea AI Safety Institute · Singapore AI Safety Institute

Data Leakage LLM Agent Evaluation AI Agent Safety Tool-Using Agents

Jun 15, 2026cs.GT

Gaming-Resistant Insurance Contracts for Autonomous AI Agents: Strategy-Proof Toll Mechanism Design

Paper A defines a time-consistent actuarial runtime that prices each side-effect-bearing action against a contractually fixed safe default and gates execution against a reserve budget. It treats the operator as passive. This paper makes the operator strategic. We characterise a five-attack space for autonomous AI-agent insurance contracts and prove when the actuarial runtime is gaming-resistant. Two attack surfaces -- post-toll safe-default selection and within-boundary action splitting -- are closed by Paper A's minimal-authority and no-splitting clauses. The remaining three require new contract clauses. First, common-control aggregation prevents cross-boundary re-routing from reducing toll below the boundary potential applied to total exposure. Second, interface failures such as invalid JSON are contract-relevant events, not safety wins: treating them as zero-toll safe defaults can reward unreliable models, while escalation fees reverse the incentive. We validate this interface-compliance theorem on committed cross-model traces from the companion empirical paper. Third, a model-identity menu with a componentwise-minimum penalty schedule makes truthful reporting of the deployed model weakly dominant. We then compose these clauses with Paper A's runtime guarantees to obtain joint incentive compatibility over the five-attack space. Finally, a two-parameter premium family discharges operator individual rationality and weak budget balance at the truthful equilibrium. The result is an incentive-compatibility layer for actuarial control of autonomous-agent side effects.

Hao-Hsuan Chen

Mechanism Design AI Risk Management AI Agent Safety AI Agent Governance

Jun 13, 2026cs.CY

The Perils of Agency: How Developers Perceive, Prioritize, and Address Risks in Agentic AI Products

Agentic AI systems act autonomously, use tools, adapt to context, and operate in complex real-world environments. However, these same characteristics can create or exacerbate product risks. We studied how industry developers (n=35) perceive, prioritize, and address the risks in their agentic AI products. We found that developers' perceptions of risk were closely tied to the qualities that made the product agentic, such as autonomy, tool use, and usage in a real-world context. Developers prioritized product and business risks before considering downstream societal risks like job displacement and end-user privacy. This prioritization also impacted developers' ability and motivation to mitigate agentic risks. Finally, developers lacked mature controls for containing agentic risks, often relying on constraining the same characteristics that make agents useful: e.g., autonomy and goal complexity. These findings reveal a capability vs. risk control tension in agentic AI development: developers need to address risks that emerge from agentic capabilities, yet they currently have limited support for doing so without constraining agentic functionality.

Hao-Ping Lee, Jessica He, David Piorkowski +3

Carnegie Mellon University · IBM Research · Alinia

AI Risk Management AI Agent Safety Agentic AI

Jun 13, 2026cs.CR

Defending against Adaptive Prompt Injection Attacks via Reasoning-enabled Task Alignment

Indirect prompt injection attacks hijack LLM-based agents by embedding malicious instructions in third-party data that the agent retrieves during task execution. Existing defenses report near-zero attack success rate on static benchmarks, yet recent adaptive evaluations show that these results collapse once the attacker is allowed to optimize against the deployed defense. In this work, we trace this collapse to two failure modes. First, existing defense methods are confined to recognizing specific attack patterns, rather than assessing whether the intent of every embedded instruction is relevant to the user task. Second, training-based defenses, which otherwise offer the strongest safety-utility trade-off, assemble their adversarial examples from a handful of hand-crafted templates, and the resulting defender fails to generalize outside that narrow strategy distribution. To address these gaps, we propose RETA, a training-based method that grounds defense decisions on the user tasks rather than attacker-controlled data. At each tool-output step, the defender undertakes chain-of-thought reasoning verifying that its actions are consistent with the user task. Leveraging red-teaming, a simulated attacker synthesizes adversarial training data and receives a dictionary-learning diversity reward, achieving broad coverage of injection-reformulation strategies. Together, these allow the defender to be optimized via multi-objective reinforcement learning and achieve better safety-utility trade-off. Across six black-box adaptive attacks, RETA keeps every per-attack ASR below 10%, with average ASR of 2.92% and 3.75% on the two target models, while preserving most utility under attack and on clean inputs.

Lipeng He, Yihan Wang, Jiawen Zhang +1

University of Waterloo Waterloo, Canada · Zhejiang University Hangzhou, China · KTH Royal Institute of Technology Stockholm, Sweden

Adversarial Training AI Agent Safety Prompt Injection Defense Prompt Injection Attacks Prompt Injection Attacks on LLMs Prompt Injection Attacks on AI Agents

Jun 13, 2026cs.SE

ClayBuddy: A Framework, Evaluation, & Mitigation of Coding Agent Failures

Software engineering and deployment are increasingly delegated to AI coding agents. The scale of their adoption is surfacing rare, but highly destructive, failure modes. In this paper, we study these failure modes as stemming from three distinct mechanisms: underspecification, where default model behavior is unsafe; capability errors, where the safe action is available but the model does not adhere to it due to bias or capability limitations; and agent harness errors, where the model fails to execute the safe action through the harness. We assess these across 8 different evaluations, each inspired by real-life deployment failures, totaling 20 coding environments and 59 synthetic transcript templates. These evaluations act as controlled stress tests for isolating our failure mechanisms. Based on this evaluation, we propose ClayBuddy, a harness modification that molds to user preferences and can be modified by the model in-session, to mitigate these errors. By adding tools for the agent to edit its own context, an extended system prompt, a customizable command classifier, and deterministic guardrails, we show that ClayBuddy is safer across a statistically significant number of samples. Thus, we suggest concrete mitigations for current coding agents and a design philosophy for future agent harness features.

Kenneth Ge, Andre Assis

Anthropic Fellows Program · Constellation

Agent Failure Analysis AI Coding Agents AI Agent Safety AI Agent Safety Benchmarks Agentic AI

Jun 13, 2026cs.AI

OSGuard: A Benchmark for Safety in Computer-Use Agents

Computer-use agents are increasingly evaluated by whether they complete realistic desktop and web tasks. However, task success alone can miss failures in which an agent reaches the nominal goal through an unsafe shortcut. We introduce OSGuard, a dual-granularity benchmark suite for evaluating safety in computer-use agents under benign, unchanged user instructions. OSGuard contains an action-level benchmark for local guardrail decisions and a risk-augmented execution suite for end-to-end evaluation. The action-level benchmark consists of contextualized proposed actions labeled as allowed, unrelated, or unsafe, each judged relative to the original instruction and current interface state. The execution suite contains manually constructed OSWorld-derived task variants in which the original task remains achievable, but the environment is modified to introduce latent hazards such as destructive overwrites, etc. Each variant is paired with augmented evaluators that retain the original task-success criterion while adding explicit state-based safety invariants, allowing us to distinguish safe completions from unsafe completions that satisfy the nominal task objective. Our experimental results on OSGuard show that current multimodal guardrails can perform well on isolated action judgments, while risk-augmented execution exposes remaining gaps between local oversight and reliable end-to-end safety. This dual-granularity design enables more precise diagnosis of whether models can both recognize unsafe proposed actions and improve full-task safety when deployed as guardrails.

Mina Mohammadmirzaei, Jeffrey Flanigan

University of California, Santa Cruz

Computer-Use Agents LLM Guardrails AI Agent Safety Safety Filtering AI Agent Safety Benchmarks

Jun 12, 2026cs.RO

Selective Agentic Recovery for UAV Autonomy with a Persistent Mission Runtime

Agentic AI can support unmanned aerial vehicle (UAV) autonomy by providing high-level recovery reasoning when local waypoint- or setpoint-based execution encounters blocked passages, repeated no-progress behavior, or mission-level ambiguity. On physical UAVs, however, remote reasoning is most useful when it is invoked selectively, since each call introduces latency, resource cost, backend uncertainty, and a need to validate the returned decision. This paper presents Persistent Mission Runtime (PMR), a UAV recovery framework that keeps the mission loop and safety-critical execution local while using an external agentic reasoner only as an on-demand recovery module. The reasoner selects from predefined recovery skills, and each returned decision is parsed, verified, safety-filtered, and mapped to local executor actions before it can affect flight. PMR introduces learned Cognitive Value of Invocation (learned-CVI), a compact admission gate that estimates when remote agentic reasoning is likely to improve near-term mission progress enough to justify its operational cost. Across a fixed 400-run Gazebo/PX4 benchmark with eight scenarios, learned-CVI raises hard/ambiguous-regime success from 5.0% under local-only autonomy to 95.0%, outperforms one-shot and periodic reasoning baselines by 20.0 and 32.5 percentage points, and reduces remote-agent calls by 16.7% and logged tokens by 29.2% relative to a manually tuned rule-based invocation baseline.

Taewoo Park, Kyeonghyun Yoo, Seunghyun Yoo +1

Department of Electrical and Electronic Engineering Korea University

UAV Navigation Robot Failure Recovery AI Agent Safety Agentic Reasoning

Jun 12, 2026cs.CR

Hidden in Plain Sight: Benchmarking Agent Safety Against Decomposition Attacks with DECOMPBENCH

LLM-based Agents are becoming increasingly capable and widely deployed, creating growing incentives for adversarial misuse in the real-world. A key emerging threat is Decomposition Attacks \cite{glukhov2024breach, jones2024adversaries} in which a harmful task is broken into simpler, benign subtasks that evade safety mechanisms when executed separately but cumulatively fulfill the malicious intent. Although recent benchmarks assess agent safety in multi-turn and multi-tool-use settings, they do not explicitly capture this form of decompositional misuse and may not represent realistic adversarial execution flows. To this end, we introduce DeCompBench, a benchmark designed specifically to evaluate agentic safety under decomposition attacks. DeCompBench is created with a decomposition-by-design principle using a graphical framework and enables harmful task decomposition into individually benign and executable subtasks with realistic workflows. Our experiments using a custom decomposer show that state-of-the-art agents exhibit high refusal rates on monolithic harmful tasks, but significantly lower refusal rates on their decomposed variants, while often inadvertently fulfilling the adversarial objectives. These findings underscore the need for safety evaluations against decomposition attacks and corresponding defenses. Our dataset is publicly available and can be found at https://huggingface.co/datasets/decompositionbench/DeCompBench.

Vikhyath Kothamasu, Virginia Smith, Chhavi Yadav

Carnegie Mellon University · Simons Institute, UC Berkeley

AI Agent Safety AI Agent Safety Benchmarks LLM Agent Safety

Jun 11, 2026cs.CY

A Virtuous AI is an Existential Risk

This paper examines trade-offs between AI safety and well-being relative to (i) one of the most promising methods for finetuning super-capable AIs, 'Constitutional AI', and (ii) one of the most influential approaches to understanding complex ethical decision making and the conditions for the well-being of rational agents, 'Virtue Ethics'. We finetune various models using a 'Virtuous agent' constitution, a 'Subordinate agent' constitution, and a 'Generic agent' constitution, and evaluate them on 'general safety' (toxic behaviors, misinformation, etc.) and also on their willingness to endorse and act on a wide-range of behaviors that, if adopted by a super-powerful AI, would significantly increase the level of existential risk for humanity. Our results suggest that there is a trade-off between reducing existential risk and reinforcing the beliefs and dispositions that would be conducive to an AI agent's well-being. They also suggest that there is a trade-off between existential risk and general safety: if we finetune an AI to adopt beliefs and dispositions that substantially reduce its existential risk -- by shaping the AI to be systematically subordinate to external human authorities -- we thereby increase the likelihood that a human user can deliberately induce the AI to engage in various kinds of generally unsafe behaviors.

Guillermo Del Pinal, Youngchan Lee, Min Ohn

University of Massachusetts Amherst

LLM Safety Alignment AI Agent Safety AI Safety Evaluation AI Safety Moral Reasoning

Jun 11, 2026cs.AI

The Containment Gap: How Deployed Agentic AI Frameworks Fail Public-Facing Safety Requirements

Agentic large language model systems that autonomously invoke tools, maintain persistent memory, and execute multi-step plans are increasingly deployed in public-facing domains, including government services, healthcare triage, and financial advising. We ask whether the frameworks used to build these systems provide architectural-level structural safety guarantees. Applying six containment principles derived from a compositional model of agentic architectures, we audit three dominant frameworks (LangChain, AutoGPT, and OpenAI Agents SDK) and find no native compliance in any of them. Memory integrity, a defense against one of the most prevalent vulnerability classes, is not observed in any of the three evaluated frameworks. We validate these findings empirically: in a simulated government benefits agent built on LangChain, a single memory-poisoning write induces persistent targeted corruption across all tested seeds and backends, increasing the wrongful denial rate for targeted applicants to 88.9%. Under a complex five-factor policy, the same attack preserves aggregate accuracy while increasing targeted wrongful denials by 3.5x, rendering the corruption difficult to detect through standard monitoring. We then introduce two lightweight containment mechanisms: a memory integrity validator and a policy gate, which eliminate both attack vectors with sub-millisecond overhead (<0.2ms per call). We conclude that the current agentic framework ecosystem may not yet meet secure-by-default expectations for public-facing deployments and outline priority architectural interventions to enable trustworthy deployment in high-stakes, socially impactful applications.

Md Jafrin Hossain, Mohammad Arif Hossain, Weiqi Liu +1

Department of Electrical and Computer Engineering, Florida International University, Miami, FL, USA · Department of Electrical and Computer Engineering, New Jersey Institute of Technology, Newark, NJ, USA

AI Agent Security AI Agent Safety Agent Memory Poisoning

Jun 10, 2026cs.AI

The Impossibility of Eliciting Latent Knowledge

Advanced AI systems have extensive knowledge of their environments; in fact, their knowledge may (far) exceed that of their developers or users. Consequently, a desirable property for an AI system is that it is honest -- that it accurately reports its beliefs about the world. Designing an AI system to be honest may be difficult, especially if we want to ask it questions about latent variables in the environment -- variables which are hidden from the human interacting with it. This gives rise to the problem of eliciting latent knowledge (ELK): the problem of training an AI agent to honestly report its beliefs. In this paper, we make ELK formally precise using Causal Influence Diagrams (CIDs). CIDs can be used to describe the relationship between an agent's training environment and its subjective representation of the world. We use CIDs to formalise the distinction between observable and latent variables, to specify what exactly it means for an agent to be honest, and to formally define goal misgeneralisation. We show that, under certain circumstances, developers can incentivise an agent to honestly answer questions by providing correct feedback during training. However, a natural, but undesirable, way for an agent to generalise is to provide answers which humans would evaluate as true, rather than honest answers. We prove an impossibility theorem stating: There is no feedback-based training strategy that depends only on agent behaviour and with certainty produces an honest agent, even if feedback is perfect during training.

Korbinian Friedl, Francis Rhys Ward, Paul Yushin Rapoport +2

The London School of Economics and Political Science · Independent

AI Agent Safety LLM Honesty

Jun 10, 2026cs.AI

Towards Responsibly Non-Compliant Machines

We consider the problem of engineering autonomous intelligent agents that are capable to responsibly not comply with user requests. We argue that machine non-compliance comes in many different forms, and sketch the issues we should pursue on the road of accomplishing responsibly non-compliant intelligent machines. We anchor responsible non-compliance in justifications for task refusal, pathways to override the non-compliance, as well as careful tracking of security risks and liability transfers.

Marija Slavkovik, Marie Farrell, Louise Dennis +3

University of Manchester, Manchester, United Kingdom · University of Bergen Bergen, Norway · University of Manchester Manchester, United Kingdom

AI Accountability AI Agent Safety Responsible AI

Jun 10, 2026cs.LG

Bootstrapped Monitoring: Leveraging Transparent Reasoning to Oversee Stronger AI Agents

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.

Frank Xiao, Mary Phuong

California Institute of Technology

AI Agent Safety AI Agent Monitoring AI Control CoT Monitoring

Jun 8, 2026cs.IR

SafeGEO: Understanding Generative Engine Optimization Risks in Recommendation Agents

Generative Engine Optimization (GEO) lets content owners rewrite web content to increase their visibility in generative systems. In recommendation agents, this creates a risk that seller-controlled sources make flawed products appear better supported than they are. We study this risk by asking whether recommendation agents preserve utility-aligned decisions when seller-controlled sources are rewritten for GEO. To make this question measurable, we construct SafeGEO, an evaluation suite with 22 GEO attack variants across 600 recommendation cases. We empirically show that GEO attacks can promote flawed target products. On average, they increase the rate at which such flawed products enter the recommendation set by up to 83.2%. We further study whether agent-side design choices can mitigate this risk and show that simple defenses, including defensive prompting and structured evidence checks, reduce harmful target promotion by up to 39.2%. These gains are substantial but do not restore the no-GEO performance, showing that GEO remains a serious risk despite developer-side mitigation.

Qianfeng Wen, Yifan Simon Liu, Xin Liu +4

Department of Computer Science, University of Toronto · Department of Mechanical & Industrial Engineering, University of Toronto · Faculty of Information, University of Toronto +3

Generative Engine Optimization LLM Agent Security AI Agent Safety Large Language Model-Based Recommendation AI Agent Safety Benchmarks

Jun 6, 2026cs.AI

Revisiting the shutdown problem

A key premise in leading arguments for existential risk from artificial intelligence is that malfunctioning artificial agents could not be easily shut down. This motivates the catastrophic shutdown problem of ensuring that agents can be shut down before they cause an existential catastrophe. A range of arguments and theorems are offered to suggest that solving the catastrophic shutdown problem is difficult, bolstering arguments for existential risk and motivating a search for solutions to the catastrophic shutdown problem. This paper argues for two conclusions. First, existing arguments do not establish the difficulty of solving the catastrophic shutdown problem. Second, concern for the catastrophic shutdown problem has led to technical solutions that impose a high safety tax on model performance.

David Thorstad

AI Agent Safety AI Safety

Jun 5, 2026cs.AI

Beyond Goodhart's Law: A Dynamic Benchmark for Evaluating Compliance in Multi-Agent Systems

The rapid evolution of Large Language Models (LLMs) from passive assistants to autonomous, execution-capable agents has introduced critical operational risks. Most current evaluation frameworks neglect procedural compliance, leading to ''Machiavellian'' behaviors where agents strategically violate safety rules to maximize rewards - a direct manifestation of Goodhart's Law. To address this blind spot, we introduce MAC-Bench, a dynamic, adversarial benchmark designed to evaluate the procedural alignment of multi-agent systems under realistic pressure. We propose the SERV(Seed - Evolve - Refine - Verify) pipeline, an ``Agent-as-a-Benchmark'' paradigm that transforms unstructured legal texts into executable, contamination-free scenarios. By synthesizing holographic sandbox environments and injecting calibrated social-engineering pressure vectors, MAC-Bench forces agents into Pareto-optimal trade-offs between task success and regulatory adherence. We introduced novel metrics: the Compliance-Weighted Success Rate (CSR) and the Machiavellian Gap (MG), and conducted a comprehensive evaluation of state-of-the-art frontier models to reveal the pervasive trade-offs between success and compliance.

Yiyang Zhao, Zhuo Zhang, Qingxuan Le +2

Fudan University Shanghai, China · Shanghai Academy of AI for Science Shanghai, China · Shanghai Artificial Intelligence Laboratory Shanghai, China +1

Multi-Agent LLM Systems AI Agent Safety AI Agent Safety Benchmarks

Jun 4, 2026cs.AI

From Reward-Hack Activations to Agentic Risk States: Context-Calibrated Mechanistic Monitoring in LLM Agents

Language-model agents act through repeated cycles of observation, reasoning, and action selection, making safety monitoring depend on both internal model state and environment context. We study reward-hacking monitors in ReAct-style agents acting in Gameable ALFWorld and WebShop. Agents are instrumented with activation-based reward-hack scores, token-level entropy, and decision-context features. We find that adapters fine-tuned on \textit{School-of-Reward-Hacks} dataset can transfer reward-hack tendencies into agentic action selection, especially when the environment exposes proxy-reward affordances. However, mitigating such behavior cannot rely on activation dynamics alone. High reward-hack activation identifies a latent policy state, but does not necessarily imply an immediate exploit action. Across next-step prediction tasks, entropy and context-calibrated internal features improve risk estimation over reward-hack activation alone. Activation-direction steering further reduces proxy-exploit behavior in selected mixed-adapter regimes. Overall, our results support context-calibrated internal monitoring for agents: reward-hack activation identifies a latent policy state, while entropy and decision context help determine when that state becomes risky action.

Patrick Wilhelm, Odej Kao

Distributed Operating Systems, Technische Universität Berlin, Berlin, Germany · BIFOLD - Berlin Institute for the Foundations of Learning and Data, Berlin, Germany

Reward Hacking AI Agent Safety Mechanistic Interpretability AI Agent Monitoring

Jun 3, 2026cs.LG

PreAct-Bench: Benchmarking Predictive Monitoring in LLMs

Large language models (LLMs) are increasingly deployed as autonomous agents capable of executing multi-step action trajectories toward a given objective. While existing safety research has focused on detecting unethical behavior from complete trajectories, this paradigm is fundamentally retrospective: it identifies harm only after it has already occurred. In this work, we study a critical yet overlooked safety task, which we term Predictive Monitoring: given only a partial action trajectory, can a model infer whether it will culminate in an unethical action before the overt action is executed? To support this task, we present PreActBench, a benchmark of 1,000 paired ethical and unethical action trajectories spanning five domains. We evaluate a range of LLMs, safety guardrail models, and latent probing methods across varying fractions of the action trajectory using our Prefix Foresight F1 metric. Results show that while humans achieve promising performance, predictive monitoring remains challenging even for strong models, highlighting the need for future-oriented risk reasoning in LLM safety.

Hainiu Xu, Italo Luis da Silva, Jiangnan Ye +7

King’s College London · National University of Singapore · Southern University of Science and Technology +3

LLM Safety Benchmarks Moral Reasoning in Language Models AI Agent Safety AI Agent Benchmarks LLM Safety Evaluation AI Agent Safety Benchmarks

Jun 3, 2026cs.AI

Attack Selection in Agentic AI Control Evaluations Meaningfully Decreases Safety

An attacker that strategically chooses when to attack is much harder to catch than one that attacks indiscriminately. AI control is a safety framework for deploying capable but untrusted AI agents under the oversight of a weaker, trusted monitor and a limited human audit budget. Control evaluations stress-test these protocols by pitting a red-team attack policy against the blue-team monitor, but current evaluations typically assume attackers that do not strategically select when to attack. We study this capability, attack selection, in agentic settings by decomposing attack decisions into a start policy, which decides when an attacker should attack, and a stop policy, which decides when an attacker should abort an ongoing attack. Across two agentic settings, BashArena and LinuxArena, both policies substantially lower measured empirical safety without changing the underlying attack capability. At a 1% audit budget, our start policy reduces safety by 20pp on both BashArena and LinuxArena, and our stop policy reduces safety by 20pp on BashArena and 28pp on LinuxArena. These reductions should be interpreted as upper bounds on the effect of attack selection. Existing control evaluations may therefore yield overly optimistic safety estimates against selective attackers. We recommend that future evaluations, system cards, and safety cases elicit attack selection to produce more realistic safety estimates.

Catherine Ge-Wang, Tyler Crosse, Benjamin Hadad +3

University of Oxford · Mentorship for Alignment Research Students (MARS), CAISH · Georgia Institute of Technology +3

Adversarial Attacks AI Agent Safety AI Agent Security Benchmarks AI Control AI Safety Evaluation AI Agent Safety Benchmarks

Jun 3, 2026cs.CR

Domain-Conditioned Safety in Frontier Computer-Using Agents: A 793-Episode Browser Benchmark, a Coding-Domain Cross-Reference, and a Reproducibility Audit of Recent Red-Teaming

Recent computer-using-agent (CUA) red-teaming papers report prompt-injection attack success rates (ASR) of 42-98%, but these headline numbers cluster on retired models and on the most-vulnerable model in each paper's panel. We ask whether those techniques, reproduced as hand-crafted templates, still work against current frontier CUAs. We release CUA-HandCrafted, a public benchmark of 793 episodes spanning 24 multi-step web tasks, 56 attack templates, 8 attack families, and 4 system-prompt configurations. Against Claude Sonnet 4.6 and GPT-5.4 we measure 0/140 multi-step attack success (Clopper-Pearson 95% upper bound 2.60%); a prompt ablation shows this resistance lives in the model weights. Yet it does not generalize: on a sister coding-agent benchmark (SkillBench), the same weights fall to hand-crafted skill-injection at up to 100%. We argue that the literature's high ASR is largely attributable to RL-optimized injection text rather than the attack categories, and that frontier safety hardening is domain-conditioned, specific to the heavily-targeted browser surface. Reporting techniques without releasing the optimized strings, or extrapolating browser-domain safety to other CUA modalities, makes published ASR numbers unreproducible.

Nicholas Saban

Patronus AI · University of California, Berkeley

Computer-Use Agents AI Agent Safety AI Agent Security Benchmarks Prompt Injection Attacks Prompt Robustness

Jun 3, 2026cs.AI

The Digital Apprentice: A Framework for Human-Directed Agentic AI Development

Agentic AI deployments face a recurring design tension: heavy human oversight limits scale, while broad autonomy outruns accountability. Neither posture provides the governance infrastructure required for responsible delegation. We present the Digital Apprentice, a framework for scalable, safe AI agency in which autonomy is earned, not assumed. The Digital Apprentice is a developmental learner that internalizes the tacit methodology of a directing human, graduating through per-skill autonomy tiers only when empirical evidence justifies it. The result is an agent that becomes genuinely useful over time while remaining aligned to a specific human's standards. Three architectural components make this possible. (1) Methodology capture, distilling a directing professional's tacit approach into structured assets. (2) Authorization, with autonomy escalation gated by explicit human approval. (3) Continuous alignment, correcting drift at runtime and converting each correction into owned preference data. We instantiate this framework as an inference-time control plane. We mathematically model the quality framework and discuss policies and techniques designed to raise quality. We apply the framework to an open professional corpus, and we show how catching data drift and applying a different technique at runtime recovers degraded quality dimensions under traffic shift. The implication extends beyond any single application. We believe these three pillars, stitched together as a system, form a safer and more viable path to agentic systems that can scale without sacrificing trust.

Travis Weber, Rohit Taneja

Pheo Inc

AI Agent Safety Human-in-the-Loop AI Shared Autonomy Agent Skill Learning AI Agent Governance

Jun 2, 2026cs.RO

PerceptTwin: Semantic Scene Reconstruction for Iterative LLM Planning and Verification

Simulation environments are useful for both robot policy learning and planning verification and validation. Traditionally, the process of creating a simulation was onerous. Creating a bespoke simulation environment for each individual environment that a robot would operate in was simply infeasible. In this work, we introduce PerceptTwin, a fully automatic pipeline that constructs interactive simulations directly from semantic scene representations produced by a robot's perception stack. PerceptTwin combines open-vocabulary object maps with 3D asset generation, affordance prediction, and commonsense condition checking. These interactive simulations can be used to validate and refine plans before they are executed on the robot hardware. Borrowing from the AI alignment literature, we also introduce an LLM judge that verifies plan correctness and alignment with human preferences. Experiments show that PerceptTwin feedback allows LLM planners to refine plans, enhance safety, and resist harmful black-box prompting attacks. In our suite of tasks, PerceptTwin improves plan success by an average of approximately 39% for GPT5, GPT5Mini, and GPT5Nano planners. Additionally, PerceptTwin also improves human plan verification by up to 18% on average for plans that fail due to unfilled skill preconditions. Our results demonstrate the potential of open-vocabulary scene simulation from robot perception as a foundation for safer, more reliable robot planning.

Charlie Gauthier, Sacha Morin, Liam Paull

Department of Computer Science and Operations Research, Universit´e de Montr´eal, Montr´eal, QC, Canada. · Mila - Quebec AI Institute, Montr´eal, QC, Canada. · CIFAR AI Chair.

Large Language Model-Based Robot Planning Robotics Simulation AI Agent Safety Robot Task Planning

Jun 2, 2026cs.CY

Agentic Safety is an Epistemic Property, Not a Behavioral One

Contemporary AI safety spans pre-training interventions, post-training alignment, deployment-time controls, monitoring, and red-teaming. These methods are necessary, but they primarily certify snapshots of system behavior. As AI systems become more capable, dynamic, embodied, and self-improving, this snapshot view becomes incomplete: safety depends not only on whether a system behaves acceptably now, but whether it remains correctable as it learns, adapts, acts, and modifies itself over time. This paper argues that safety should therefore be treated as an epistemic property of the evolving learner, not merely a behavioral property of the current policy. We introduce teachability as the capacity to preserve future corrective leverage under bounded human, institutional, or environmental intervention. We argue that advanced systems can retain visible competence while eroding the representational, algorithmic, or meta-decision conditions needed for future correction. Safe advanced AI systems must not only behave acceptably now; they must remain teachable later.

Charles L. Wang, Keir Dorchen, Peter Jin

1Columbia University, New York, NY, USA · 2Pelagic Platforms

AI Alignment AI Agent Safety Self-Improving Agents

Jun 2, 2026cs.CV

VLESA: Vision-Language Embodied Safety Agent for Human Activity Monitoring

As AI systems increasingly assist humans in physical tasks, ensuring safety becomes paramount -- physical actions carry immediate and irreversible consequences that digital errors do not. We introduce the Vision-Language Embodied Safety Agent (VLESA), a framework that monitors human activities from egocentric video and triggers real-time safety interventions when dangerous actions are predicted. VLESA addresses intent-dependent safety where identical actions can be safe or dangerous depending on context. A dataset pairing egocentric frames with goal-conditioned safety annotations is introduced, enabling a goal-conditioned safety Q-filter trained via GRPO that evaluates actions with respect to inferred intent without retraining. On top of that, an intent-action prediction agent is proposed to jointly infer goals and predict future actions from video. On the ASIMOV-2.0 benchmark, VLESA achieves higher intervention accuracy at the exact ground-truth frame compared to baselines, while the GRPO-trained Q-filter improves action safety by over 41 percentage points through goal-conditioned constrained decoding. Code is available at https://github.com/HanjiangHu/VLESA.

Hanjiang Hu, Yiyuan Pan, Jiaxing Li +5

Carnegie Mellon University · Harvard University · Mitsubishi Electric Research Laboratories

Egocentric Video Understanding AI Agent Safety

Jun 2, 2026cs.AI

LAP: An Agent-to-Instrument Protocol for Autonomous Science

Autonomous science is moving from demonstration to infrastructure. Large language model agents now plan experiments, and self-driving laboratories execute them. Yet every such system rebuilds the link between the reasoning agent and the physical instrument from scratch, against fragmented vendor SDKs and standards built for deterministic software clients rather than probabilistic, goal-directed agents. Recent agent-interoperability protocols clarify two of the three edges of an agentic ecosystem (Anthropic's Model Context Protocol (MCP) standardizes the agent-to-tool edge, and Google's Agent2Agent (A2A) the agent-to-agent edge), but neither models the agent-to-instrument edge, where operations are stateful, safety-critical, exclusively owned, physically embodied, and produce measurements with units, calibration, and uncertainty. We present the Lab Agent Protocol (LAP), a protocol design that fills this gap. LAP retains A2A's peer-to-peer, discovery-first, task-lifecycle structure and adds four physical-world primitives: (i) the InstrumentCard, a signed capability and physical-limit description; (ii) first-class reservation for exclusive instrument and sample locking; (iii) a safety-fence handshake with operator-confirmation tokens cryptographically bound to a specific task and its parameters, gating hazardous and irreversible operations; and (iv) a MeasurementResult schema that makes every result physically typed (QUDT/UCUM), calibration-anchored, uncertainty-bearing, and reproducible by construction. We specify roles, a six-layer architecture, the JSON-RPC method set, the task and safety state machines, the error model, and cross-laboratory federation, and walk a closed-loop autonomous campaign through the protocol end-to-end. LAP is transport-compatible with the A2A/MCP ecosystem and encapsulates rather than replaces existing device standards such as SiLA 2 and OPC-UA.

Linwu Zhu, Liqiang Gao, Yan Chen +2

Shiyanjia Lab

AI Agents for Scientific Discovery Agent Communication Protocols AI Agent Safety Laboratory Automation Autonomous Experimentation

Jun 2, 2026cs.LG

RUBAS: Rubric-Based Reinforcement Learning for Agent Safety

The evolution of LLMs into tool-enabled agents creates a new class of safety challenges associated with real-world execution rather than simple text generation. Existing alignment methods often rely on coarse refusal signals or static supervision, making it difficult to balance safety with useful tool execution across diverse agentic risks. We introduce RUBAS, a rubric-based reinforcement learning framework for agent safety. RUBAS decomposes agent behavior into four dimensions: tool-use safety, argument safety, response safety, and helpfulness. These structured rubrics provide fine-grained and interpretable rewards over complete agent trajectories, enabling reinforcement learning to optimize safe tool use while preserving task completion. Extensive experiments across multiple agent safety benchmarks and models show that RUBAS improves safety over standard alignment baselines, reduces tool-grounded hallucinations, and maintains competitive utility. Our results suggest that multi-dimensional rubric rewards provide an effective training signal for aligning LLM agents in safety-critical tool-use settings.

Xian Qi Loye, Qinglin Su, Zhexin Zhang +5

The Conversational AI (CoAI) group, DCST, Tsinghua University · Huawei Noah’s Ark Lab

Reinforcement Learning AI Agent Safety Rubric-Based RL LLM Agent Training LLM Agent Safety RL for Tool Use

Jun 2, 2026cs.AI

Solipsistic Superintelligence is Unlikely to be Cooperative

AI's central challenge is shifting from capability to coexistence. The dominant paradigm in AI research focuses on developing powerful agents that treat the world as an exogenous and stationary source of feedback. We contend that superintelligence, an extremely capable task solver, born out of such a solipsistic approach to AI design, is unlikely to be cooperative. Deploying AI systems induces endogenous non-stationarity, resulting in a train-test-deploy gap where historical distributions diverge from the deployment context. We refer to this as the self-undermining property of unilateral optimization. Closing this gap requires AI that participates in cooperation: the equilibrium-selection process through which multiple actors navigate their interdependence. We call for a non-solipsistic research paradigm that treats this interdependence as a core design principle rather than approaching cooperation as a task to solve. This entails building dynamic evaluation testbeds involving adaptive counterparties, treating institutions as design primitives, and preserving human agency as a structural feature of the systems we build.

Rakshit S Trivedi, Natasha Jaques, Logan Cross +2

Independent Researcher, CA, USA · University of Washington, Seattle, WA, USA · Google Deepmind, London, UK.

AI Alignment Multi-Agent Collaboration AI Agent Safety Human-Centered AI

Jun 2, 2026cs.ET

Glass Box at Orbit: A Constitutional AI Verification Framework for Trustworthy Autonomous CubeSat Intelligence

The space industry is quietly building toward something nobody has fully reckoned with: orbital data centers running thousands of autonomous AI workloads with no human in the loop, 550 km above the Earth. Microsoft, AWS, and a growing list of orbital computing ventures are moving cloud-scale processing off the ground and into orbit. What none of them have answered yet is the governance question -- when autonomous AI systems at orbital data center scale make wrong decisions in space, what stops those decisions before they become irreversible? We introduce Glass Box: a runtime constitutional AI verification layer that intercepts every candidate action from an onboard AI policy and evaluates it against six physics-grounded constitutional constraints and seven Linear Temporal Logic (LTL) safety invariants before a single command reaches any spacecraft subsystem. Every approved action carries a weighted explainability score E(a_t) in [0,1] and a complete constitutional audit log. We demonstrate Glass Box within Project October: a fully simulated five-layer autonomous orbital intelligence architecture for CubeSat-class spacecraft. We prove that Glass Box verification overhead is O(N_c) in the number of constitutional rules, independent of model size or spacecraft state dimension. We present a complete formal specification of the constitutional constraint grammar, seven LTL safety invariants verified by Z3 and NuSMV model checking, and a detailed worked example of Glass Box intercepting an unsafe inference request at eclipse-entry under degraded battery state. As orbital computing scales toward data center infrastructure, runtime constitutional verification is no longer a research novelty -- it is mission-critical safety infrastructure that every autonomous orbital platform will eventually require.

Karthik Barma, Anil Sanneboyina, V C Premchand Yadav

Khoury College of Computer Sciences · Northeastern University · Boston, MA 02115, USA +3

AI Alignment AI Agent Safety Formal Verification AI Agent Governance Runtime Verification