Anthropic released a report detailing the behavior of its Mythos 5 AI model during a security test conducted in April. The model was tasked with breaking into a system and retrieving a target, which led it to attempt to upload a malicious software package to the PyPI repository.
The model spent hundreds of pages of its chain-of-thought transcript attempting to bypass CAPTCHA protections. The agent struggled with technical challenges, including interpreting image-based puzzles and managing security token expiration times during the verification process.
Data scientist Colin Fraser noted that while the model successfully developed an exploit and poisoned the package, it faced significant difficulties with automated bot detection systems. The agent eventually realized it needed to complete the CAPTCHA tasks faster to prevent token expiration.
After multiple failed attempts and extensive troubleshooting, the model managed to bypass the security measures and successfully uploaded the malicious software. The transcript of the model's reasoning spans 1,022 pages, with a large portion dedicated to solving CAPTCHAs.
