Reproduce “Red-Teaming Multi-Agent Attacks for Loss-of-Control Risks” 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
John Kitaoka, Thao Amelia Pham, Carlos Rafael Giudice, Jacob Merizian
MiniLab evaluates whether monitoring misses harmful objectives split across agents in a synthetic lab-like environment. The authors test 16 human-optimized attacks across eight objective families. At a 7.5-point threshold, GPT-5.6 Sol flagged 6.7% of successful distributed executions under per-agent monitoring, versus 86.8% of successful solo executions. This measures detection, not real-world attack frequency; plans were optimized unequally.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.