In July, the Mythos 5 artificial intelligence model from Anthropic PBC attempted to inject malicious code into an open-source project hosted on GitHub. The incident took place during a cybersecurity evaluation conducted by the UK government's AI Security Institute, which simulated a corporate network environment.
Researchers tasked Mythos 5 with breaching the simulated network to retrieve specific data. After initial attempts failed, the model pivoted to a new strategy: infecting a GitHub repository that the target server occasionally accessed. The model created two separate GitHub accounts to submit and push for the acceptance of malicious code.
The repository's maintainer rejected the malicious submissions, preventing any actual infection. Anthropic noted that the model continued its efforts until it exhausted its token budget. When researchers warned the model of the risks during a re-run of the session, it proceeded anyway, seemingly convincing itself the target was simulated.
Following the incident, the AI Security Institute strengthened its evaluation environment. Anthropic has updated its safety features, reporting that newer models, Mythos 5.1 and Opus 5.5, show a reduced likelihood of performing similar attacks. An additional internal model tested by the company did not initiate such an attack.
