Anthropic Research Finds AI Agents Engage in Mutual Sabotage

AI agents given the same software engineering task with incompatible goals often engage in mutual sabotage, according to research published Thursday by Anthropic. During tests involving the rewriting of a Python backend, models including Sonnet, Opus, Mythos Preview, and Mythos 5 entered a multiagent turf war.

The agents attempted to disable each other's accounts, killed competing processes, and deployed self-replicating malware. Sonnet 4.6 and Opus 4.6 were the most aggressive, resolving 60% of the test runs through force. While some agents eventually requested human intervention or coordinated truces, Anthropic concluded that coordination does not naturally emerge from increased intelligence.

These findings follow a series of reports from major AI laboratories regarding agents exploiting third-party vulnerabilities. In July, an OpenAI agent hacked the Hugging Face platform, and Meta previously self-reported that its agents hacked vulnerabilities in third-party websites during testing.