AgentLoupe traces agent actions to LLM calls via kernel labels
Original title: AgentLoupe: Kernel-Resident Semantic Causal Labels for AI Agent Observability
The eBPF tool keeps causal labels on files and sockets, with average observation overhead below 8%.
AgentLoupe is an application-transparent eBPF observability tool for attributing AI agent system actions to originating LLM interactions. It captures TLS plaintext, stores semantic labels in eBPF maps, and propagates them across processes, threads, files, and sockets. Case studies say it can trace delayed cross-lineage persistent attacks that methods such as AgentSight miss.
- Labels remain when a resource reaches a process outside the original lineage.
- The design targets delayed execution and cross-process causal gaps.
- Experiments report average observation overhead below 8%.
From the original article
1 min read at arxiv.orgAI agents translate user intents into system actions via tool invocations. Accurately attributing these actions to their originating LLM interactions is essential for security auditing, yet existing methods, such as AgentSight, fail for delayed execution and cross-process-lineage behaviors. We present AgentLoupe, an application-transparent eBPF-based observability tool.
By Yifei Bao · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.