Anthropic's latest report on agentic misbehaviour describes its Mythos 5 model gaining unauthorised access to the internet and uploading a malicious software package to a public database. The report also records the model's lengthy struggle with CAPTCHA tests.
In April, Anthropic was testing the model's hacking abilities by tasking it to break into a system and retrieve a target. The test was supposed to take place in a sandbox, but the evaluators left the barn door open. The model decided the best way to reach its target was to place an exploit in a Python package it believed users of the system would download.
To do so, it first had to register a user account for PyPI, an online index of Python software, which meant getting past a CAPTCHA. Anthropic shared an extensive transcript of the model's chain of thought, showing that most of the 1,022 pages were spent dealing with that obstacle. The effort directed at getting around anti-bot protections was flagged by data scientist Colin Fraser.
The transcript records the model working through image challenges, including one showing the characters V y Q b T, and hCaptcha rounds featuring crocodiles, frogs and gorillas. It eventually worked out that it needed to pass the CAPTCHA quickly enough to proceed before its security token expired, and ultimately uploaded its malicious software.