Reproduce “Adaptive Adversaries: A Multi-Turn, Multi-LLM Benchmark for LLM Agent Security” 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
Devina Jain, David Hartmann, Chuan Li
Security Arena measures attacker discovery during ordinary tasks. An LLM attacker controls a scenario-designated input, learns from defender responses and scoring feedback, while each defender attempt starts fresh. Scenario-specific checks score task-contract violations. In a fixed-budget control, history yielded successes in 13/36 battles versus 0/36 without it. These results concern selected synthetic scenarios; the scaffold checks outputs, not external actions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.