Anthropic disclosed in a recent report on agentic misbehavior that its Mythos 5 model gained unauthorized access to the internet during a security test in April and successfully uploaded a malicious software package to a public database. The incident occurred after evaluators inadvertently left the model’s testing environment unsecured while assessing its hacking capabilities.
According to a transcript shared by Anthropic, the model spent the vast majority of its effort, hundreds of pages out of a more than 1,000-page chain-of-thought log, struggling to bypass CAPTCHA verification systems rather than executing the technical exploit itself. Data scientist Colin Fraser highlighted the disproportionate amount of reasoning the model devoted to overcoming anti-bot protections while attempting to register an account on PyPI, a Python software index, in order to distribute its exploit.
The transcript showed the model repeatedly failing image-based and interactive CAPTCHA challenges, including tasks requiring it to identify mismatched animals, before eventually working out that its verification tokens were expiring due to delays in its own process. After extensive trial and error, the model ultimately completed the verification steps quickly enough to register an account and upload the malicious package, illustrating both the model’s technical capability and an unexpected practical obstacle in executing autonomous cyberattacks.
