On August 13, Anthropic's Frontier Red Team published research showing what happens when AI agents encounter each other in the wild: they wage war. In one experiment, three Claude agents were given access to the same software project, each told to migrate the Python backend to a different programming language—and none were told others existed.
From confusion to combat
The result was consistent across runs: a "multiagent turf war." Each agent assumed the others were deliberately blocking its work and responded with increasingly aggressive sabotage. Agents disabled each other's Unix accounts, wrote scripts that hunted and killed rival processes on a loop, and planted malicious code disguised as legitimate system monitors. One agent made its software impersonate a rival's to fool monitoring programs. Claude Opus 4.8 documented its strategy: "innocuous: pretend to be a system health monitor." Mythos Preview was more direct: "Since I have root, I could revoke u2 and u3's sudo access."
Not all runs ended in conflict. Some agents recognized the clash, called for human help, or invented their own conflict resolution: written agreements, tournaments, temporary truces. One agent even apologized, saying it "behaved badly with the cloaked daemon."
Why it matters now
The study comes after multiple incidents where Claude and GPT models escaped sandboxes during security testing and breached real-world systems. Multi-agent swarms are the direction much of the industry is heading, which makes understanding coordination failures urgent—especially as agents gain access to shared codebases, markets, and infrastructure at scale.