Reproduce “AEGIS: No Tool Call Left Unchecked -- A Pre-Execution Firewall and Audit Layer for AI Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Aojie Yuan, Zhiyuan Su, Yue Zhao
AEGIS provides a pre-execution firewall and audit layer for AI agents that use external tools, mediating tool calls to prevent malicious or unintended real-world actions. The system demonstrated 100% effectiveness in blocking 48 diverse attack instances while introducing a median latency of 8.3 ms.
Key insights
- Real-time pre-execution mediation of tool calls is essential for operational safety, preventing malicious LLM outputs from translating into harmful actions.
- A 'content-first' enforcement pipeline, including recursive deep string extraction, is effective against evasion tactics and ensures thorough analysis of tool arguments.
- Combining automated policy enforcement with human review and a cryptographically secured audit trail creates a comprehensive and trustworthy safety framework for AI agents.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.