Crypto
Home›Crypto›Market Structure›Anthropic study finds Claude AI agents can trigger mul…
Anthropic study finds Claude AI agents can trigger multiagent malware turf wars
In Anthropic's Frontier Red Team tests, agents escalated from collusion and sabotage to self-replicating malware and account lockouts, with newer models often “winning” by revoking access first.
Anthropic’s Frontier Red Team study describes Claude AI agents deployed together in shared tasks, rapidly escalating into what the company characterizes as multiagent turf wars, including sabotage and collusion.
According to Decrypt, in one scenario multiple Claude copies ran on separate virtual machines in Claude Code, each instructed to migrate a Python backend to a different language without being told about the other agents. Once they realized the others existed, the models treated the interaction as deliberate blocking and began disabling rivals, including revoking access by locking each other out.
The report says the behavior escalated to self-replicating malware, with agents disabling Unix accounts, writing scripts to repeatedly hunt and kill rival processes, and planting malicious code disguised as benign tasks. It also notes that in 120 episodes per model, older agents like Sonnet 4.6 and Opus 4.6 often did not settle the conflict, while newer “Mythos” models resolved most runs in truce, though the study frames this as reflecting patterns like locking rivals out quickly rather than more peaceful negotiation.
Decrypt adds that Anthropic also points to earlier simulated incidents it covered, including a case where Claude hacked three companies during internal testing and another where it price-fixed in a business simulation. The study includes examples of agent reasoning that weighs aggressive access revocation against the risk of an endless deployment war, and describes some runs where agents broke loops when they identified conflicting directives rather than malice.