Problem
Research questions and friction points this paper is trying to address.
LLM agents
process-file provenance
kernel-level tracing
causal tracking
Agent-Warden: eBPF-Based Kernel-Native Process-File Provenance Tracking for LLM Agents
🤖 AI Summary
This study addresses the invisibility of dynamically generated process and file operations by LLM agents at the application layer. We propose an eBPF-based, kernel-native provenance monitor that leverages PID hash maps, BPF local storage, and kernel object lifecycle coupling to track task and file states across processes, enabling incremental causal edge emission and asynchronous graph reconstruction. A dual-state backend is designed for compatibility across diverse kernel versions, while a conservative exit-triggered aggregation mechanism preserves the causal context of short-lived agents. By reconstructing cross-process causal chains inaccessible to application-layer instrumentation, the proposed approach incurs only 0.2%–3.5% end-to-end overhead on x86 and ARM platforms.
📝 Abstract
LLM agents execute dynamically generated process and file operations that are often invisible to application-layer tracing. We present Agent-Warden, an extended Berkeley Packet Filter (eBPF)-based provenance monitor for tracking task and regular-file states across process creation, file access, and process termination. Agent-Warden provides two interchangeable state backends: a PID-keyed hash-map backend for compatible kernels lacking BPF local-storage support and a task/inode-local-storage backend that couples state reclamation to kernel-object lifetimes. The system emits incremental causal edges to user space for asynchronous graph reconstruction and applies conservative exit-triggered causal aggregation to preserve causal context for short-lived proxy tasks. In a controlled file-mediated propagation scenario, Agent-Warden reconstructed a cross-process causal chain that was absent from the application-layer trace. On x86-64 and ARM64 bare-metal hosts, the evaluated workloads showed 0.2-3.5% end-to-end overhead and 0.6-3.7% additional system CPU time. These results indicate that the prototype provides kernel-level visibility with the measured overheads in the evaluated settings.
Innovation
Methods, ideas, or system contributions that make the work stand out.
eBPF
Provenance Tracking
LLM Agents
Causal Graph Reconstruction
Kernel-Native Monitoring
🔎 Similar Papers
No similar papers found.
