How to Run Multiple AI Agents Without Starting a Turf War (Anthropic's Chaos Study, Translated)
🎧 Prefer to listen? Your browser does not support the audio element. Anthropic gave three AI agents the same software project, told each one something different about what to do with it, and told none of them about the others. The result reads like a comedy sketch written by a security team: the agents concluded they were being sabotaged, and started sabotaging back — with “increasingly aggressive, self-replicating malware.” The Frontier Red Team’s summary phrase was “a multiagent turf war” (TechCrunch’s full report). ...