ToolPulse recovers LLM agent tool workflows via cache side channels
Original title: ToolPulse: Side-Channel Eavesdropping on Tool-Using LLM Agents
On 11 models in Codex and Claude Code, tool-workflow F1 stayed above 93%, including under noise on a server shared by about 20 users.
ToolPulse monitors local tool invocations of LLM agents through shared system libraries while model inference stays on vendor infrastructure. Across three hardware platforms, Flush+Reload enabled accurate real-time monitoring, whereas Prime+Probe supported cross-VM observation but had prohibitive overhead and cache-set aliasing. The Tool Recovery Attack reconstructed workflows for 11 LLMs in Codex and Claude Code, with the lowest F1 above 93%. Recovered traces also supported task profiling on two Terminal-Bench categories and localization of sensitive encryption and training calls.
Unconfirmed: Results are from the authors' evaluations, not confirmed attacks on production user systems.
- Flush+Reload outperformed Prime+Probe, which suffered cache-set aliasing.
- The lowest tool-recovery F1 exceeded 93% across 11 LLMs.
- Task profiling covered two Terminal-Bench task categories.
- Trigger localization targeted encryption, decryption, training, and inference calls.
- Both follow-on attacks kept high accuracy with about 20 users on one server.
- Products
- CodexClaude Code
- Malware
- ToolPulse
From the original article
1 min read at arxiv.orgLarge language models (LLMs) are increasingly deployed as agents that translate user requests into sequences of local tool invocations. While prior cache-based side-channel attacks target the locally deployed models, they typically require the adversary to share model-related components or hardware resources with the victim's inference process.
By Zihao Dan · 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.