CARE (Canonicalization, Attribution, and Resolution Engine), a shell-specific, static-first verifier for individual shell commands before execution can reduce dispatch-boundary risk for LLM agents while preserving most benign workflows.
Abstract
Large Language Model (LLM) agents are increasingly used for coding and terminal automation, making shell-command dispatch a high-stakes runtime control point. We study command-level pre-execution mediation for individual shell commands produced by LLM agents under bounded path context. Existing safeguards remain limited: generic guardrails do not model shell structure in sufficient detail, always-on LLM judges are relatively costly and variable, and shell parsers do not directly prevent harmful execution. We present CARE (Canonicalization, Attribution, and Resolution Engine), a shell-specific, static-first verifier for individual shell commands before execution. CARE canonicalizes generated commands into stable verification targets, derives deterministic evidence over syntax, command semantics, path context, and provenance-backed risk patterns, and escalates only underdetermined cases to an LLM judge. This design keeps the common case fast, reproducible, and auditable while reserving neural adjudication for borderline commands. On the balanced main split, CARE reaches 85.64% F1 with a 0.91% false-positive rate at 2.32 ms mean latency. When deployed in its static enforcement profile, CARE retains 84.99% F1 at 0.34 ms and reduces realised harm on RedCode-gen to 37.33%. Across external-generalization tests and controlled Docker-sandbox execution, these profiles expose a practical trade-off between benign recovery, false-positive burden, latency, and harm reduction. Overall, command-level shell mediation can reduce dispatch-boundary risk for LLM agents while preserving most benign workflows.
AgentGuardUtil is presented, the authors' entry to CAR-bench Track~1, which treats the AI planer (LLM) as a fallible proposer inside a grounded verify-and-revise loop, and its core novelty is a runtime policy compiler.
R. Bouchekir, Damir Safin, Tomas Bueno Momcilovic· 0 citations
Large language model agents place outputs from external skills into their execution context, allowing attacker-controlled data to influence later privileged actions. Existing defenses mainly classify untrusted content or authorize proposed operations. They do not directly address how an agent's future authority should change once untrusted data enters its state. We present SkillGuard, a harness-level enforcement layer that treats this event as contamination and restricts future capabilities to disconnect the resulting state from deployer-defined forbidden states. Given sound skill summaries and policies, SkillGuard represents security-relevant transitions with a Skill Impact Graph, specifies admissible control over skill parameters via steerability signatures, and mediates invocations with an inline reference monitor. Following contamination, it computes weighted capability restrictions using binary, fractional, or fractional-flow strategies without auxiliary language-model inference. We evaluate SkillGuard on four AgentDojo suites with two backend LLMs, Gemini 2.5 Flash and Llama3.3-70B, against an LLM-only No Defense baseline and three defenses at different system layers: Spotlighting, CaMeL, and AttriGuard. We construct a compositional attack benchmark in which each attack combines observations individually insufficient to induce target violation and evaluate the same baselines on it. Under AgentDojo's Tool Knowledge attacks, SkillGuard eliminates attack success on three of four suites for both backends and reduces it to 4.8% and 14.3% on Slack. Against compositional attacks, it outperforms every baseline on Llama and matches the strongest baseline on Gemini at higher benign utility. Fractional-flow restriction preserves substantially more capabilities than binary restriction at the same attack success rate. Across both settings, SkillGuard adds no model calls or token overhead.
Wu-Jie Xiong, Rabimba Karanjai, Yang Lu et al.· 0 citations
This work argues 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.
Kai Wang, Zeming Wei, Biaojie Zeng et al.· 0 citations
Enterprise agents must follow long-horizon, conditional, safety-critical standard operating procedures (SOPs), which are compiled into executable pseudo-code and run with a program-guided stack machine that pages the active frame while an LLM performs semantic execution.
Chenglin Yu, Lichao Yin, Qingxin Fan et al.· 1 citation
Deployed large language model (LLM) agents are now being used to interface with external tools, fetch information, run code, interact with user data and help with decision making at the workflow level. Therefore, their safety issues are not only related to the underlying model, but also to tool permissions, prompt design, retrieval sources, policy constraints, evidence validity, and varying operational contexts. While Goal Structuring Notation (GSN) has become a popular means of structuring safety arguments, many GSN-based safety cases are still reliant on manual inspection and can suffer from unsupported goals, incomplete evidence paths, undischarged assumptions, circular arguments, unaddressed defeaters, or insufficient coverage of hazards. This paper introduces MCSafe-GSN-HOL, a formal assurance framework for creating machine-checkable safety cases for deployed LLM agents that overcomes these disadvantages. The proposed framework translates the structured GSN arguments into Isabelle/HOL proof obligations, and formalizes the typed GSN elements, the hazard contracts of LLM-agents, the evidence-completeness conditions, the assumption-discharge rules, the defeater-resolution rules, and the incremental re-verification rules after deployment changes. The prototype was tested with 25 safety-case fragments from five representative scenarios of LLM-agents. Results demonstrate that MCSafe-GSN-HOL has 97.3% defect detection, 97.2% evidence coverage, 96.4% hazard-control coverage, and 62.7% reduction in review time compared to manual GSN review. The mean Isabelle/HOL checking time was 2.53 s/safety case, showing that theorem-prover supported assurance can still be practical for medium scale safety-case checking. The results overall support the conclusion that formally checked GSN safety cases can help with the traceability, auditability, consistency and maintainability of assurance processes for deployed LLM-agent systems.
Aakash Abhay Yadav, Shashank Shelat, B. Hinduja et al.· International Conference on...· 0 citations
SkillEffect is presented, a checked-lowering runtime for computations with a recoverable source relation, an audited bounded implementation, and a registered output postcondition that shows that one checked-lowering architecture can enforce heterogeneous registered memory relations at Agent tool dispatch.
Yinuo Wang, Yiyu Shi· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
