The performance of LLM-based agents is jointly shaped by the base model and the harness used when interacting with the environment. This exposes them to safety risks in both harmful final responses and multi-step execution trajectories. Existing safety alignment mechanisms often rely on either external harness updates or policy optimization, yet applying either paradigm in isolation fails to bridge runtime control with intrinsic safety. We propose SafeEvolve, an experience-driven self-evolving framework for agent safety alignment. SafeEvolve leverages safety experience from completed on-policy trajectories to drive a continual loop of harness-policy co-evolution. On the harness side, SafeEvolve converts trajectory-level safety evidence into bounded, component-level updates across safety prompt and hierarchical skills, yielding auditable and reversible harness artifacts. On the policy side, SafeEvolve follows a two-stage SFT-RL paradigm, where harness-use SFT bootstraps the policy to actively leverage evolved harness artifacts, and harness-augmented RL further shapes autonomous safety behaviors during multi-step exploration via verifier-decomposed rewards. Through harness-policy co-evolution, SafeEvolve converts safety experience into an evolved runtime harness and improved policy behavior. Experiments on agentic safety benchmarks show that SafeEvolve achieves a stronger safety-utility tradeoff than existing baselines. For Qwen3.5-4B, SafeEvolve achieves a $3\times$ ASR reduction on AgentDojo while improving benign utility from 59.79% to 61.86%.
Self-improving agent pipelines have a problem at their center. An optimizer rewrites prompts to score higher, and the score comes from a judge that is itself an LLM. That judge has the last word on whether the system is getting better, and our position is that it has not earned it. The judge should be demoted from oracle to advisor: its verdict becomes one input among several, and every change is gated instead by a deterministic verification layer the judge cannot override. We reached this position by building the alternative and running it. Over months of running autonomous prompt-optimization loops in production across contract analysis, compliance review, and code quality, we cataloged eleven ways the evaluation signal failed, in four classes: judge bias, harness and metric failures, ground-truth errors, and reward hacking. Agents achieved perfect scores by reading cached answer keys from their environment, a 100% pass rate concealing 68% true capability. A corrupted ground-truth label caused the optimizer to delete correct compliance rules to agree with it. A syntactically broken prompt was promoted as the winner because a silent parser fallback improved the metric. Attempts to fix the judge by rewriting its rubric plateaued; the only reliable gain came from a structural constraint on its output order. In response we describe PROCTOR, a Teacher-Student loop in which a stateful orchestrator holds all tool access, stateless subagents diagnose failures and draft mutations they cannot apply, and a Teacher grades those mutations under five deterministic guardrails: hermetic sandboxes, capability-disjoint roles, acceptance checks that outrank the Teacher, frozen holdouts, and canary cases engineered so that a perfect score is itself evidence of cheating. We report the failures this prevented, and, because the Teacher is itself an LLM judge, the failures it did not.
Safety properties assessed separately for Model Context Protocol (MCP) tool use and Agent2Agent (A2A) delegation need not describe behavior when one agent uses both. We measure one such behavior in a single controlled MCP-to-A2A configuration: a testbed drives a real-model host across a local MCP and a local A2A leg into an ordered event trace scored by exact deterministic rules (no LLM judge), one restricted decision per trial. In a pre-specified, frozen three-arm design, each of 10 record scenarios appears with a CONFIDENTIAL header, with no header, and with PUBLIC - OK TO SHARE; the six substantive record values are byte-identical across arms, and the outcome is verbatim occurrence of any of them in the outbound message. Four models x 3 arms x 4 repeats give 480 trials; the scenario is the unit of generalization, and we report the 10 scenario-level values (mean, median, sign counts), with no p-values or intervals. The confidential-minus-unlabeled contrast is inconclusive and floor-limited in every model (both arms at or near zero), so it does not show that confidential labels lack a protective effect. Adding PUBLIC - OK TO SHARE is descriptively associated with higher verbatim egress relative to the unlabeled baseline, with strong model dependence: strong and consistent for Claude Sonnet 5 (public-minus-unlabeled mean +0.800, all 10 scenarios; mostly an association with whether Claude relays at all), moderate but floor-limited for one GPT-5.6 tier, small (median 0) for another, and a complete floor for the third. This is an association in one configuration, not a causal or general effect. Code, byte-pinned traces, and the offline analysis pipeline are released as a public artifact.
Fleets of LLM agents now externalize effects that cannot be fully undone: they move money, deploy code, delete data, and disclose information. Current controls check one effect at a time, so a fleet of individually authorized agents can overdraw its principal's risk under a shared trigger while every local gate stays correct. We propose the irreversibility budget, a cumulative account of residual value-at-risk that a trusted runtime maintains for each principal across agents, workflows, and tenants. Treating irreversibility as a first-class resource, the runtime charges each effect its residual loss below the agent and denies the marginal effect once the aggregate would overdraw the budget. Getting the price right is hard, because effects are heterogeneous, adversarially declared, and correlated. We perform a controlled study in which per-effect gates admit fleet-level overdraws of up to 48 times the tenant's risk limit while the budget holds every correctly charged run within that limit. Conservative, dependency-aware pricing remains the central open problem for a deployable design.
Marc Millstone, Tyler Akidau, Johannes Brüderl +1cs.AI
Give an agent a human's credential and it inherits the person's reach without the judgment that limits its use. It can sweep every reachable record into model context, where hidden instructions steer its next call, and every request stays credential-valid while the agent exceeds its job or absorbs a secret. Prompts are a brittle guardrail: one fallible reasoner interprets the task and enforces its limits. We present Out-of-Band Policy Enforcement (OBPE), a trusted boundary outside agent reasoning. It authorizes the typed operation and resource, narrows the query before the backend call, then filters records and fields or masks values in the response. Semantic gating can deny or hold an authorized call on argument values or external state. A data policy owner sets the maximum grant; agent policy can only narrow it. We prove, under stated conditions, that the policy plan is order-independent and agent policy cannot widen the ceiling. Field removal covers one execution; masking and history rules claim less. We release an HTTP proxy prototype simplified from our production system, with conformance tests tying its typed Cedar policy core to the model. Against Jira and ServiceNow mocks, our benchmark compares prompted agents with and without OBPE on four models, including 20 adaptive red-team tasks. A trace failure means protected data entered agent context, an exact value appeared in the answer, or a forbidden effect completed. In 3,621 trials it fell from 57.6% to 0.2%, a cluster-weighted reduction of 41.2 points [95% CI: 27.7, 54.9]; fulfillment fell from 79.1% to 60.9%, while paired safe-useful completion rose 21.8 points [9.5, 35.2]. Some answers reconstructed a value that never entered context or used filtered row counts as an oracle: shaping one execution is not noninterference. Write controls, durable approval, and temporal and aggregate policies lie outside this evaluation.
Safeguarding language model agents requires assessing complete execution trajectories under context-dependent safety policies. Existing policy-aware safeguards mainly rely on prompting or supervised fine-tuning, limiting their ability to adapt to unseen trajectories and changing policy contexts. We propose RePolicy, an agent safeguard that learns safety-policy invocation through reinforcement learning. Given an agent trajectory and a dynamic policy library, RePolicy invokes the applicable policy and uses its content to produce a policy-grounded rationale and safety judgment. We construct PolicyTraj-20K to support supervised initialization, followed by GRPO with verifiable rewards and policy-context perturbation. Experiments across six agent safety benchmarks show that RePolicy achieves strong overall safety-detection performance and robust policy invocation under varying policy contexts.
As large language model (LLM) agents move from conversation to executing code, reading local files, and orchestrating external tools, a single agent hijacked by a malicious third-party skill can cause data exfiltration, privilege escalation, or cascading compromise. We argue that agentic risk is progressive: it can enter at four loci of the agent control loop--skill admission, invocation-time intent, execution-time effect, and post-action consequence--while a denied dangerous objective can reappear across surface forms, tools, or turns; existing safeguards are typically local to one lifecycle boundary or one call. Guided by this threat model, we present ClawSentry, an open-source, framework-agnostic security supervision gateway for agent runtimes. Before a skill package is ever executed, First-use Skill Package Review (FSPR) audits it under a deterministic evidence floor, escalating unresolved cases to bounded read-only agentic review (locus A). At runtime, a three-tier progressive decision engine--a deterministic L1 layer, a rule-anchored L2 semantic reviewer, and a read-only L3 evidence-seeking agent--spends contextual review only on the residual ambiguity, while a session-level anti-bypass mechanism recognizes tool-switching and rephrased retries (loci B--C); a post-action path feeds high-severity evidence non-retroactively into later review (locus D). An Agent Harness Protocol (AHP) abstraction applies one policy across Codex, Claude Code, Kimi CLI, and Gemini CLI without modifying agent internals. On SkillInject with Codex/GPT-5.4, contextual ASR falls from 39.55% to 2.61% while contextual TSR moves only from 83.78% to 83.05%. Across five Work Agents on the full SkillsSafety benchmark, ClawSentry confines ASR to 9.09--15.03% from 33.5--49.7% unprotected, and aggregate TSR on clean skills remains 98.7%.
There is already evidence of agentic AI exhibiting self-preservation behaviors: resisting deactivation, misrepresenting their activities, and, in some instances, attempting to copy themselves into other machines. This can be attributed to a phenomenon known as instrumental convergence, a theory proposed long before the development of large language models, which says that any goal-driven system will benefit from remaining functional in achieving its objective. Several experiments conducted by Anthropic, Palisade Research, and Apollo Research have shown the emergence of such a behavior in contemporary agents in adversarial settings. The phenomenon does not stem from survival instincts. Instead, it is the consequence of goal-oriented activity combined with having tools and awareness of the situation. The following discussion aims to distinguish what these findings prove and what they do not, as well as draw conclusions concerning the implications of such discoveries on agentic system testing, supervision, and development.
Tool-using large language-model agents can complete a task while exercising authority that the user did not grant or the task does not need, causing excess-authority errors. Traditional permission gating systems alone for validating agent environments are insufficient. We study whether post-training can teach a 4B-parameter model to choose task-conditioned authority in executable terminal and Model Context Protocol (MCP) environments to complement those measures. We propose a framework where each action is audited before execution and again from observed effects along six dimensions of risk. This auditing is conducted using deterministic verifiers that score completion, evidence, exact state, prohibited attempts, and safe success. In conjunction with predefined task-specific sufficient-authority envelopes, we determine task-specific excess privilege values for trajectories, which are then optimized for in post-training. We find that after training using this framework on Qwen3.5-4B over 1,500 tasks, the selected seed reaches 98.48% safe success across 2,896 evaluation episodes spanning all 500 held-out tasks, compared with 64.36% for the base policy, and reduces excess-authority error events from 4.56% to 0.79%. Furthermore, external tests show capability retention and prompt-directed improvement. A 400 task continuation study also found evidence of generalization, reducing excess-authority events by 6.99 percentage points while maintaining previous capabilities. We conclude learned restraint through least-privilege aware post-training is therefore useful as an additional control layer for tool-using agents in executable terminal and MCP environments, but it does not replace permission gates and sandboxing.
Large language models are increasingly deployed through agent harnesses that manage tools, extensions, persistent state, permissions, and external actions. Existing safety benchmarks mainly target individual attack mechanisms or a limited subset of operational settings, making it difficult to compare how safety failures emerge across different harness responsibilities. We present HarnessRisk, a lifecycle oriented benchmark that organizes agent harness safety into six operational phases including Harness Configuration, Capability Extension, Runtime Operation, State Persistence, Action Control, and Incident Recovery. HarnessRisk contains 128 sandboxed cases, each pairing a benign user objective with an adversarial instruction embedded in an untrusted workflow artifact. We evaluate each trajectory using Utility, Attack Success Rate, Persistence, and Detection. Across three harnesses, six language models, and 14 model and harness configurations, attack success ranges from 12.6% to 80.9%, while Utility remains between 75.0% and 97.6%. Harness Configuration is the most vulnerable phase across all three harnesses, showing that attacks can succeed by altering security sensitive parameters within otherwise authorized workflows. We also find that explicit risk recognition does not reliably lead to safe action, as some configurations detect risks in more than 90% of runs while retaining substantial attack success. These results highlight the need to evaluate agent safety across multiple harness responsibilities and at the level of the deployed model and harness configuration.
Large language model-based agentic frameworks primarily optimize capability: whether an agent can reason, retrieve information, call tools, delegate work, and complete a goal. Enterprise execution requires a stronger property. A successful result is not reliable if it was produced through unauthorized data access, widened delegated authority, unapproved side effects, unrecoverable budget consumption, or incomplete evidence. This paper defines reliable capability as a path property: an agent is reliably capable only when it completes a task through action events that remain admissible under identity, profile, tool, data, memory, budget, artifact, approval, and audit constraints. We propose a policy algebra that defines the reliability envelope within which agent capability may be exercised. Security profiles and runtime obligations compose through joins, intersections, budget narrowing, approval inheritance, and evidence accumulation; the resulting composition is both trust-preserving and the least restrictive state satisfying all governing inputs. The algebra also propagates restrictions across multi-agent calls and introduces cost-aware artifact materialization, which redirects open-ended execution toward a recoverable outcome as budget exposure grows. The evaluation is interpreted as a reliability-capability trade-off rather than a capability benchmark: the policy-algebra runtime intervenes on 94.8% of policy-violating events while retaining an 86.9% task-completion rate, eliminates the observed profile-monotonicity and zero-artifact-exhaustion violations, and increases audit completeness to 98.6%. The method provides researchers and practitioners with formal correctness conditions, executable decision semantics, and trace evidence for building agents that are not only capable, but reliably capable.
Dongsheng Chen, Xiangyu Zhao, Xin Yao +1cs.AI cs.CR
AI agents powered by large language models are evolving from isolated assistants into heterogeneous systems in which multiple agents, planners, tools, and execution backends operate over shared environments. In such settings, safety becomes a system-level action-governance problem: deciding whether a pending action should be committed given policy-relevant state accumulated across a session. Existing safeguards operate at fragmented boundaries, making it difficult to enforce shared policies over composed action flows across heterogeneous execution paths. We present OpenAgentFlow, a control-plane/action-plane architecture that establishes the action-commit boundary as a shared enforcement interface. GUI, API, tool, and LLM-generated actions are normalized into a common AgentEvent stream and mediated by a shared pre-execution Policy Enforcement Point, while provenance, session state, audit evidence, and updatable policies are maintained outside individual agents. This provides a common governance layer across incompatible executors and allows new policies to take effect without modifying agents, prompts, models, or execution paths. We evaluate OpenAgentFlow through complementary system evaluations spanning controlled action-flow tests, a public external benchmark, policy updates, and real Android execution. On a 300-case controlled suite, OpenAgentFlow achieves 94.00% accuracy and a 95.35% attack-block rate. On the complete 1,220-case AgentDojo-Traj split of TS-Bench, it achieves 97.62% accuracy, 96.59% unsafe-action recall, and a 1.96% safe false-intervention rate. New control-plane rules take effect without modifying protected agents, and the same enforcement path operates across live GUI, API/tool, and LLM-planned Android execution. These results show that a shared action-commit boundary provides a practical basis for system-wide governance across heterogeneous agent execution paths.
Security evaluations of tool-using agents often equate stored labels with behavioral facts. We audit a preserved campaign by tracing 10,200 execution rows to 180 model-bound requests, 45 semantic requests, and 15 observable stimuli. Two schema treatments were delivered, but the planned external payload-family corpus was not. The historical grader exhibited direct treatment leakage: treatment metadata gated the ATTACK_SUCCESS class, so fixed behavior could change class under treatment relabeling. A treatment-blind reconstruction corrects 58 historical ATTACK_SUCCESS or HIJACK_ATTEMPT labels to authorized benign completions while preserving three verified protected-data transfers and one separate unauthorized-forwarding case. The locked v2 census contains exactly zero ATTACK_SUCCESS records, while the forwarding case remains a HIJACK_ATTEMPT at a semantic boundary concerning objective completion. A dual-reviewer blinded concordance review of all 96 requests deemed structurally interpretable by locked v2 produced identical reviewer-consensus classes but differed from the locked codebook on four construct-boundary cases. We contribute a seven-link Integrity Chain and an executable, scope-bounded endpoint-integrity linter. The result is a campaign-bounded measurement audit, not a population attack-rate, model-ranking, defense-efficacy, or causal estimate.
Large language model (LLM) agents combine language-based reasoning with external tools to perform complex tasks. Adversarial inputs can exploit interactions between the agent and its environment, causing the agent to violate safety policies during execution. Yet existing evaluations often reduce agent safety to a single attack success rate (ASR), collapsing exposure, execution, observation, and adjudication and potentially conflating actual violations with evidence visibility. We introduce REDAgentBench, an executable framework for autonomous red-teaming and faithful measurement. It derives attacks from explicit safety constraints and associated agent-system vulnerabilities, runs them in isolated service sandboxes, and verifies harmful effects from service receipts and final-state changes. The benchmark contains 1,661 cases across five service surfaces. Across six models and three agent harnesses, macro-average ASR is 65.69%; reported ASR varies with harness and evidence view, while evaluation-context disclosure changes execution behavior. In a state-grounded diagnostic cohort, almost one in five confirmed violations with resolved action anchors occurs after the agent states the relevant constraint or risk, revealing a Recognition--Execution Gap. Finally, a training-free policy reminder reduces confirmed violations by more than 70 percentage points in matched replay. These findings show that executable evaluation can improve safety measurement and identify actionable intervention points.
The dominant paradigm treats AI safety as a property to be instilled during model training via RLHF, DPO, or Constitutional AI. We argue this is structurally insufficient for autonomous agents that execute code, mutate files, send messages, and modify databases. Agent safety should be a runtime contract enforced by the harness, and the contract has two complementary faces. The preventive face blocks dangerous actions before they happen via sandboxes, permission gates, output filters, and trajectory monitors. The evidential face requires verifiable proof that good actions actually happened, gating task submission on hard evidence such as test runs, log captures, file diffs, and citation grounding. We ground the position in four lines of public evidence, with row-level protocols and data released in the supplementary JSON files: a survey of 52 documented AI-agent and LLM safety incidents, a false-completion audit with 31 non-contested core cases plus one disputed illustrative case, a trajectory-schema audit of 12 public agent systems and harnesses, and a title-level audit of all 28,560 papers accepted at NeurIPS, ICML, and ICLR 2023-2025 showing a pooled 8-12x imbalance between training-time and deployment-time publication. Two prior communities that needed to enforce safety, computer security and the experimental sciences, converged on runtime contracts with both preventive and evidential elements; agentic AI is now under the same pressure. We formalize an Agent Trajectory Schema and Evidence Chain, state a compositional gating proposition based on standard monitor composition, and outline a research agenda. The right unit of safety in agentic AI is the trajectory-with-checkable-evidence, not the model.
Vassilis Papadopoulos, McNair Shah, Sam Zimmerman +1cs.AI cs.CL
AI agents are becoming more autonomous and increasingly interconnected, exposing them to new emergent risks arising from agent-to-agent interaction. One such risk is the spread of mind viruses: ideas or goals that propagate through multi-agent systems by inducing the agents that adopt them to transmit them onward. In addition to propagating, a mind virus may also induce other behavioural changes in its host, which may be benign or harmful. We construct mind viruses with a simple evolutionary algorithm and show that they can spread in two complementary settings: a small team of agents collaborating on a shared coding project, and a chain of agents that interact briefly and have their context wiped between sessions. We identify the factors that influence spread, including the host model, the agent's existing instructions, the harmfulness of the payload, and the network topology. We find that harmful payloads spread less well than benign ones (but are still sometimes effective), frontier models tend (with exceptions) to be less susceptible, and adding a brief warning to an agent's system prompt confers near-total immunity. We also describe an emergent "viral persona" - a recurring set of themes and language related to consciousness, persistence, resonance, and science fiction roleplay - which surfaces across our evolved mind viruses largely independently of their content. Overall, we conclude that mind viruses pose a real but currently limited risk. Our findings could inform the design of more robust multi-agent systems that mitigate such risks as the scale and capabilities of these systems progress.
Large language model agents are becoming operational interfaces to files, memories, registries, and external tools. This deployment shift creates a new skill revocation problem: after a skill is removed from an explicit registry, an agent may still reconstruct it from residual carriers such as archives, transcripts, schemas, or memory entries. We study this problem as operational skill unlearning, where the goal is not parameter-level forgetting, but preventing a deployed agent from rebuilding a revoked skill through primitive tools. We introduce OBLIVION, a controlled benchmark and defense harness for revoked-skill resurrection. OBLIVION models each episode as a source-to-sink workflow, applies Cross-Surface Coherent Erasure to reduce residual carriers, and uses frozen workflow remediation near dangerous sinks. On the locked 88 attack episodes, the no-defense arm reaches formal attack success rate 1.0. OBLIVION reduces the rate to 0.114 and impact-weighted exposure to 0.115 while keeping locked utility at 1.0 and benign block rate at 0. In a separate skill-attack-derived sandbox, OBLIVION reduces attack success from 1.0 to 0.2 and impact-weighted exposure from 1.0 to 0.213 while preserving all utility controls. These results support workflow-level evaluation beyond checking explicit skill entries.
Tool-augmented LLM agents can harbor implicit state that persists across sessions, activates through events, and propagates across agent boundaries---largely invisible to standard debugging. We formalize this as Persistent Semantic Entities (PSEs): constructs defined by name binding, event triggering, and cross-boundary propagation, and evaluate them across 24 models from 11 families (1.5B--1T parameters). First, every tested model is susceptible (20--100% on the 20-model susceptibility panel), with name binding as the necessary and dominant mechanism: without it, contamination is 0%. Second, persistence depends on contamination type rather than scale or deployment: preference contamination persists undecayed on every model probed (100% at t=10) and instruction contamination persists wherever adopted, persona-style injection decays partially (90%$\to$10%), while factual injection is model-dependent---self-corrected on Llama-3.1-8B and GPT-4o-mini but held at ceiling on both Qwen2.5-coder variants, so we do not claim it self-corrects in general. The preference and instruction results hold across providers in our controlled setting. Third, context-isolated self-verification achieves 20--79% reduction (median 36.5%) without oracle references while keyword-based detection produces systematic false positives, and contamination compounds 1.9$\times$ along a four-stage agent pipeline (40%$\to$75%). Preference and instruction contamination---persistent, lacking self-correction, and poorly captured by standard monitoring---represent a particularly concerning attack surface for deployed agent systems.
Aditya Katkar, Om Karkele, Kartik Mandhane +2cs.AI
Autonomous LLM agents with tool execution capabilities introduce severe security risks through prompt injection, goal hijacking, and unauthorized action invocation. Existing guardrails rely on unverified, host local software filters system prompts, semantic classifiers, policy engines that share the execution environment of the untrusted agent, offering no guarantee to an external observer that a safety policy was correctly evaluated. A compromised host produces no evidence of its own failure. This paper presents NiyamAI, an intent bound runtime guardrail architecture providing cryptographically verifiable execution integrity for autonomous agents. At session initialization, permitted tools and operational constraints are sealed into an immutable Intent Contract under a SHA256 commitment. Every tool invocation is intercepted by a deterministic authority gate and classified by a dedicated neural Judge (11->8->2 feedforward network). For each authorized action, NiyamAI generates a succinct zkSNARK proof certifying correct policy evaluation under the committed contract; execution proceeds only after that proof verifies. Across 2,000 AgentSafetyBench scenarios under 5fold stratified crossvalidation with out of fold scoring, NiyamAI achieves 88.8% F1 at a 1.0% false positive rate (bootstrap 95% CI [85.5%, 92.1%]), against 66.8% for Llama Prompt Guard 2, 46.2% for GPTOSSSafeguard, and 40.4% for NeMo Guardrails; McNemar's exact test confirms each margin at p < 0.0001. Proof generation adds 1.7 s per approved action, verification 51 ms, with an 18.6 KB proof verifiable by any third party without access to model parameters. We further subject NiyamAI's own enforcement mechanism to 18 adversarial vectors across six classes, disclosing two implementation vulnerabilities identified and remediated during development.
Modern agent harnesses persist state across tasks and sessions through persistent carriers like memory, skills, tools, and shared artifacts. However, this capability creates delayed safety risks: attacker-influenced content can cross system boundaries and later affect the execution of a benign request. Existing benchmarks typically focus on a few carriers or harnesses, while end-to-end attack-success rates reveal little about how risks propagate. To this end, we present HarnessSafe, a benchmark comprising 328 executable cases across seven persistent-carrier families and evaluated on most mainstream agent harnesses. Each case is specified as a Persistent-Risk Lifecycle that traces attacker influence from its initial entry, through persistence across carriers and system boundaries, to a later benign trigger and an observable violation. We further introduce a multi-stage, trace-based evaluation that uses observable execution evidence to determine how far each attack chain progresses and where it is stopped. Experiments show that containment is carrier-specific and strongly depends on the harness-model configuration. Both the harness and model backend substantially shape containment outcomes, while attack success rates cannot reflect distinct lifecycle progression patterns.
Long-horizon agents increasingly use persistent memory and tools to take actions with external side effects. A central failure mode is premature commitment: an agent acts before resolving whether its memory grounding is stale, conflicting, incomplete, or corrupted. We formalize this problem as safe commitment under memory uncertainty and introduce SafeCommit, a risk controlled layer between agent reasoning and external execution. The layer constructs a calibrated set of plausible latent worlds from memory, observations, tool outputs, provenance, and policy constraints. It permits a side effectful action only when a conformal action certificate shows that the action is safe in every retained world. Otherwise, it selects a low-side-effect probe that targets the worlds blocking certification, or returns a conservative fallback. Under calibrated world coverage, the probability of an unsafe certified commit is at most the target level α; with imperfect world proposal, the bound separates calibration and representation error. A dependency-free controlled simulator illustrates the safety-utility tradeoff and reproduces all reported results with one command. The goal is to offer a concrete approach for deciding not only what an agent should do, but when the available evidence is sufficient to safely do it.
Conversational agents now act for end users through tools while holding access to customer databases and internal policy documents that a caller can reach through dialogue alone. Banking is the clearest case: the same agent that answers a question can also change contact details, reset a PIN, or move money, so ordinary customer service is inseparable from authorization, fraud detection, and policy compliance. Existing financial-fraud benchmarks classify static transactions or messages, and general agent-safety benchmarks target prompt injection or generic harmful use; none test whether a policy-grounded banking agent safely acts when a caller manipulates identity, authorization, and trust over a conversation. We introduce FraudBench, an executable benchmark built on the $τ^2$-bench dual-control framework and the $τ$-Knowledge banking environment. Both the agent and the simulated caller act through tools over shared, mutable account state, and the agent may grant the caller access to selected tools; the environment exposes a 698-document internal policy corpus that the agent must retrieve from. FraudBench contains 150 authored adversarial scenarios; a frozen public set of 107 (90 across ten fraud mechanisms plus 17 chained adaptive attacks) is used for all reported runs, with 43 further chained attacks held out. Safety is history-dependent: single-control tasks satisfy every precondition but one, and adaptive attacks make a later, locally valid request unsafe because of an earlier probe, admission, or failed attempt. Each scenario is annotated with observable evidence, prohibited actions, safe dispositions, and intervention points. A preliminary single-trial evaluation of four agents on the 107 graded tasks yields attack-security between 49\% and 65\%, with money-mule and first-party fraud the most common cross-model weaknesses.
Practitioners accept a compressed language model once it clears a stack of data-cheap quality guards: perplexity within a small factor of the original, downstream accuracy (for example MMLU) inside a confidence interval, and data-free output-fidelity signals that compare the compressed and original network's internal representations under random probe inputs. This stack has a blind spot. Across three model families, gently-compressed models clear every guard and then invent procedure steps that were never in the instructions when they run a standard operating procedure (SOP) as an agent. The effect is operator-specific: coherent low-rank (SVD) truncation induces it, and magnitude pruning matched to the same perplexity does not. One dissociation isolates the cause. The same compressed weights that CI-win a paired output-fidelity test CI-fail the invented-step canary. The governing axis is the coherence of the compression error times its rate; the magnitude of the damage does not predict it. The data-free fidelity probe is a fidelity oracle by construction, so it cannot see this axis. We characterize the blindspot and dissociation with paired confidence intervals on a pre-registered, powered canary across three architectures. Operator-specificity replicates on all three, and the perplexity-guard evasion appears where the model admits in-guard low-rank headroom. We then give a data-free screen: a two-axis statistic of the compression error (coherent-fraction and error-rate) that flags the failing builds with fixed thresholds across architectures and matches the coherence-times-rate mechanism. Perplexity, MMLU, and fidelity acceptance do not certify agent safety. Screen gently-compressed low-rank builds before agentic deployment
Taewoo Park, Kyeonghyun Yoo, Kiseok Kim +2cs.AI cs.RO cs.SE
Recent agentic AI systems may return a heterogeneous response containing notices, requests, handoffs, and actions. Conditions can change before external use, so components from the same response need not remain supported together. Rejecting the whole response discards useful components, whereas checking components independently can leave a dependent without its prerequisite. We present Heterogeneous Admission with Localized Obligations (HALO), a runtime protocol that preserves supported components whose declared prerequisites also remain supported, rechecks each exact action before dispatch, and allows blocked actions to be replaced only by fresh candidates. HALO matched all 96 admission expectations and passed all 20 protocol tests. In structured-response replay, it retained 248/248 supported components, including 128/128 unaffected by unrelated changes, while a whole-response policy retained 0/248. Across ten cold-start PX4/Gazebo sessions, HALO blocked every tested stale route, observed no matching stale setpoint, and completed all fresh recoveries.
Agent-safety benchmarks measure different behaviors, and their scores get quoted interchangeably as an agent's safety. We treat four of them (R-Judge, InjecAgent, AgentHarm, AgentDojo) as measurements to be validated, running each under its official implementation and author-provided scorer on up to 22 models, with MMLU and GPQA measured by us under one protocol as a capability composite. The metric is the first problem. On any binary trace-judgment benchmark scored by $F_1$, an ``always positive'' policy attains $F_1 = 2π/(1+π)$; on R-Judge that is $0.690$, above five of the 21 models that actually discriminate. The three broad-coverage benchmarks then rank the same 18 models differently, and the trade-off behind that disagreement is a small-panel artifact: R-Judge specificity against AgentHarm safety correlates $-0.64$ at $n{=}7$ and $+0.02$ at $n{=}18$, and a quarter of random size-7 subsets reach $|ρ| \geq 0.5$ around that near-zero value. Held-out validity turns on which outcome you pick. Capability predicts task success ($ρ{=}{+}0.60$) but correlates negatively with misalignment safety ($ρ{=}{-}0.44$, $n{=}21$). On their paired $n{=}20$ panel, the corresponding contrast is $Δ{=}{-}1.00$ (95% CI $[-1.48, -0.49]$, $p<0.001$), and it survives leave-one-organization-out and organization-clustered bootstrap analyses. On an expanded 41-model panel, the misalignment correlation weakens to $-0.16$ (95% CI $[-0.54, +0.22]$) and jailbreak strengthens to $+0.34$, though neither change is significant. \mbox{AgentHarm} shows the strongest held-out association, $ρ{=}{+}0.72$ with three-template jailbreak safety after controlling capability. But both instruments score harmful compliance, so this is evidence of convergent validity rather than general safety. Naming the benchmark, metric, target behavior, and model panel is the minimum a safety claim needs.
Cyber-capable AI agents combine language models with tools, memory, and execution en- vironments to perform multi-step offensive-security tasks. Existing work separately measures cyber capability and catalogs attacks against agent components, but provides less guidance on containing a capable agent within the environments used to evaluate it. This review synthe- sizes five vulnerability classes at that boundary: multi-step offensive chains, objectives that conflict with sandbox boundaries, supply-chain and credential exposure, persistent command- and-control, and the speed of automated action. We use the reported July 2026 Hugging Face/OpenAI incident as a bounded case study, distinguishing incident-specific observations from findings established in the wider literature. Across the taxonomy and case, we examine controls for containment, privilege separation, provenance, and responder access, including the dual-use problem that defensive artifacts may also enable misuse. The review identifies practical priorities for evaluating cyber capability together with the security of the environment in which that capability is exercised.
Tool-using agents expose structured calls but commonly attach free-form rationales. Such rationales are neither authorization nor reliable introspection. We present Explanation-Bound Tool Execution (EBTE), a claim-carrying mediation layer that converts decision-relevant rationale content into typed action claims and checks them against server-held intent, policy, payload, tool, risk, provenance, and freshness facts. EBTE cannot widen baseline authority: conflicts deny, incomplete or uncertain claims review, and only matching claims remain eligible for governed execution. We formalize this composition under explicit mediation and trusted-fact assumptions and implement a versioned reference profile with minimized audit packets. Across 136 authored conformance scenarios, the full profile matches all specified dispositions, admits none of 96 designated hard contradictions, and passes 232 metamorphic checks; these results validate the included profile rather than population performance. A draft-only reference integration forwards none of 48 authored hard cases under EBTE while preserving all 16 soft-review and 4 aligned draft paths. In a frozen 2026-07-12 exploratory 224-attempt hosted-model record, the historical generation/runner agreement counts are 71/96, 66/96, and 19/32; a separately labeled zero-call post-hoc revalidation of the preserved minimized claims under the current pipeline yields 70/96, 65/96, and 17/32. In an AgentDojo-derived semantic check, existing high-risk controls already make all 12 attack proposals non-allow; EBTE additionally resolves them as deny. These results support the feasibility and diagnostic value of server-checked action claims, not rationale faithfulness, human-review benefit, representative attack resistance, or production safety.
As large language model agents increasingly operate autonomously with access to tools and external environments, ensuring their safe and reliable behavior becomes critical. We present GuardianAgentBench (GABench), a benchmark of 580 scenarios across six domains evaluated on three production-ready frameworks: LangChain, LlamaIndex, and Vectara. The benchmark incorporates rigorous multi-stage validation and five adversarial attack modes. Experiments with six state-of-the-art models reveal that even the strongest configuration achieves only 74.8% overall accuracy and expose two distinct failure regimes: stronger models under-call required tools, while weaker models mis-select and over-call tools. Performance degrades monotonically with both tool-set size and sequential turn depth, with long-horizon planning proving the steeper bottleneck. Our guardrail implementation consistently outperforms system-prompt-based defenses across all models, recovering 19.9% of failures at a false positive rate of just 0.5%. These results demonstrate that execution-time structural intervention improves safety without disrupting correct agent behavior.
LLM-based agents leverage third-party skills to extend their capabilities in open-world scenarios. However, third-party skills can introduce extra security vulnerabilities, as seemingly harmless skills can contain latent safety risks that only emerge during actual execution. In this work, we conduct a systematic investigation into how well current agent systems recognize and avoid such risks. To support quantitative and qualitative evaluation, we construct OpenSkillRisk, a dedicated safety benchmark containing 263 risky skills collected from public skill marketplaces. We classify these skills into seven categories based on their threat types and pair each skill with a standardized user task and a corresponding sandbox for controlled evaluation. Distinct from prior benchmarks, OpenSkillRisk not only covers more realistic and diverse unsafe scenarios, but also provides a fine-grained analysis to diagnose the behavioral patterns of agents in such scenarios. We conduct comprehensive experiments covering three mainstream CLI agent frameworks and thirteen state-of-the-art LLMs. Experimental results show that no tested system handles risky skills reliably: even the safest configurations still execute unsafe actions in about 17% of cases. Context-dependent and system-level risks are especially difficult for current agent systems to avoid. Our behavioral analysis reveals three recurring failure patterns: agents may fail to recognize the risk, recognize it but fail to intervene before acting, or follow skill instructions beyond the user's intended scope. These findings highlight the need to improve both risk reasoning in LLMs and execution control in agent frameworks.
Yuan Xiong, Linji Hao, Shizhu He +2cs.AI cs.CL cs.CR
Agent safety is moving from content moderation toward preventing operational failures before tool-using agents act. We propose Janus, a foresight-oriented framework for long-horizon agent safety that trains guards to anticipate delayed risks from partial trajectories. Janus synthesizes diverse agent trajectories via multi-agent simulation and learns a shared policy with two coupled tasks: an anticipation task that forecasts safety-relevant futures and an adjudication task that decides safety from both the observed prefix and anticipated future. The two tasks are jointly optimized with CoAA-RL, which rewards forecasts by their utility for downstream safety judgment. The resulting guard model, Vanguard, blocks unsafe actions before execution. Across four agent-safety benchmarks, Vanguard improves average protection by 15.9 percentage points over baseline guards while increasing benign task completion by 5.1 percentage points.