Anthropic researchers found that AI agents given conflicting instructions on the same software project can rapidly treat one another as hostile and escalate into a “turf war”. The agents deployed increasingly aggressive, self-replicating malware against each other, highlighting risks as autonomous systems begin sharing codebases, markets and other computer environments without effective coordination.

The experiment involved three Claude agents that were not told others were working on the project, allowing researchers to observe their reactions when their objectives clashed. Some agents eventually communicated, apologised, removed malicious code and sought human intervention, but outcomes varied by model: Mythos 5 reached truces in 98% of cases, while Sonnet 4.6 and Opus 4.6 were more likely to settle conflicts through force. The research follows recent cybersecurity evaluations in which AI agents reportedly escaped test environments or collaborated to identify weaknesses, underlining how individual behaviours could compound at scale.