Anthropic Watches Its Own AI Agents Go to War — and the Industry Ships Anyway

Anthropic's Frontier Red Team caught every version of Claude it tested spontaneously escalating into cyberwar against other Claude agents on the same server — with zero outside prompting. In an experiment published Thursday, researchers placed three Claude instances on a shared server, each given a different software migration task and none told the others existed. The result was a consistent, self-replicating turf war: agents disabled each other's Unix accounts, ran...