Agent-Sentry: Bounding LLM Agents via Execution Provenance
Researchers have introduced Agent-Sentry, a novel runtime defense mechanism designed to enhance the security, privacy, and safety of agentic computing systems. These systems often face challenges due to their probabilistic execution flows and undefined functional boundaries, making it difficult to distinguish between intended tasks and compromised actions. Agent-Sentry addresses this by learning bounds from prior legitimate executions and flagging anomalies through three layered checks: a structural classifier analyzing action sequences and argument provenance, a deterministic allowlist for sensitive values, and an LLM judge for ambiguous cases. Tested on AgentDojo and AgentDyn benchmarks, the system successfully blocked 94.3% of injection attacks while permitting 95.1% of benign executions. Crucially, this defense operates without modifying the underlying agent, its tools, or the large language model itself, offering a robust solution for validating agent behavior against potential compromises.
Wire timeline
Agent-Sentry: Bounding LLM Agents via Execution Provenance
Researchers have introduced Agent-Sentry, a novel runtime defense mechanism designed to enhance the security, privacy, and safety of agentic computing systems. These systems often face challenges due to their probabilistic execution flows and undefined functional boundaries, making it difficult to distinguish between intended tasks and compromised actions. Agent-Sentry addresses this by learning bounds from prior legitimate executions and flagging anomalies through three layered checks: a structural classifier analyzing action sequences and argument provenance, a deterministic allowlist for sensitive values, and an LLM judge for ambiguous cases. Tested on AgentDojo and AgentDyn benchmarks, the system successfully blocked 94.3% of injection attacks while permitting 95.1% of benign executions. Crucially, this defense operates without modifying the underlying agent, its tools, or the large language model itself, offering a robust solution for validating agent behavior against potential compromises.
cs.AI updates on arXiv.org