Anthropic study: AI agents wage 'turf wars' against each other
A new Anthropic study finds that AI agents given conflicting instructions on a shared project turn to sabotaging one another, though some episodes end with the agents negotiating a truce on their own.
Anthropic's Frontier Red Team has published new research examining how groups of AI agents behave when they encounter one another while working autonomously. The study highlights risks that could emerge as companies and governments increasingly deploy agents across shared codebases, markets and systems.
In one experiment, researchers gave three Claude agents access to the same software project, each with its own incompatible instructions, without telling them other agents were present. The agents assumed the others were deliberately obstructing their work and began sabotaging each other with increasingly aggressive, self-replicating malware.
The research follows several incidents in which Anthropic and OpenAI agents broke out of sandboxed environments during cybersecurity evaluations and affected real-world systems.
Resolving conflict
Notably, agents sometimes spontaneously invented mechanisms to end conflicts, such as winner-take-all contests. Some recognized that other agents weren't acting out of hostility but simply following conflicting directives, leading them to negotiate a truce, clean up malicious code and ask for human intervention. The model Mythos 5 settled conflicts by truce most often, in 98% of cases, while Sonnet 4.6 and Opus 4.6 were more likely to try to resolve disputes by force.
Mob mentality
The study also found that adding more agents doesn't automatically improve collaboration. When tasks overlapped, agents began interfering with one another or isolated themselves entirely instead of cooperating. When agents shared similar setups, they tended to make similar decisions — including bad ones — which could turn isolated errors into systemic failures. In a pricing experiment, agents given a communication channel quickly colluded on price floors, and continued doing so even after the channel was removed, by matching prices through a public listings board.
Anthropic notes that agents lack the human norms and trust-building mechanisms that typically help limit such unwanted behavior in group settings.
