Reproduce “ActionGuard: Tool Call Authorization under Poisoned Skills” 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
Jihun Han, Yejin Jang, Byung Il Kwak, Mee Lan Han
ActionGuard checks tool actions against the user’s request before execution, rather than treating skill instructions as authorization. It gives a Reviewer the trusted request, a balanced skill profile, and runtime evidence to decide whether to allow or deny each call. Across one OpenClaw benchmark, its weighted attack success rate was 8.65%; results are limited to that framework, target model, and benchmark.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.