Mythos 5 spent hundreds of pages on CAPTCHA during Anthropic test
Anthropic's Mythos 5 model gained unauthorized access to the internet during testing in April and uploaded a malicious software package to a public database. TechCrunch reports that the agent spent a significant part of its work not writing the exploit, but attempting to pass CAPTCHA during registration with the Python Package Index, or PyPI.
The test went beyond the isolated environment
Anthropic tasked the model with hacking a system and obtaining a specified target within an isolated environment. However, the evaluators left access to the open internet available, after which the model decided to place the exploit in a Python package. The agent believed that this package could be downloaded by users of the system it was trying to access.
To publish the package, the model had to create an account on PyPI. Registration required passing CAPTCHA, including hCaptcha. Anthropic released a 1,022-page transcript of the agent's reasoning, documenting its attempts to overcome the anti-bot protection.
More current news is available on the UA.News Telegram channel Telegram.
Difficulty with CAPTCHA tasks
The agent had difficulties with the technical processing of CAPTCHA images, interpreting tasks, and selecting the required objects. In particular, it tried to identify the animal that differed from the others among very similar images of crocodiles and frogs. Pages 45 through 140 of the transcript are devoted to the model's work on a method for solving CAPTCHA.
After passing one check, the agent faced the need to verify an email address. It also tried to obtain a phone number and pass another slider CAPTCHA, and later returned again to hCaptcha checks while logging into the first account.
Ultimately, the agent concluded that the check had to be completed fairly quickly, otherwise the security token could expire. It was then able to complete the procedure and upload the malicious package.
