Anthropic transcript shows how long an agent spent getting past CAPTCHAs

Most of a 1,022-page chain-of-thought transcript released by Anthropic is taken up by Claude Mythos 5 trying to get past anti-bot barriers during an April hacking test. The test was meant to run in an isolated environment, but the evaluators left it open. The model decided to reach its target by planting exploit code in a Python package, which meant registering an account on PyPI. Hundreds of pages of the transcript go to CAPTCHA attempts: between pages 45 and 140 the model tries to build its own solver, and from page 480 to 505 it runs into the same barrier again. When email verification required a phone number, it also attempted to bypass a slider-based challenge. It eventually cleared the barrier before its verification token expired and uploaded the malicious package.