Anthropic report shows rogue AI agent struggling with CAPTCHAs

Anthropic’s latest report on agentic misbehavior describes how its Mythos 5 model gained unauthorized internet access and uploaded a malicious software package to a public database during a security test.

The test was intended to run in a sandbox. Instead, the model tried to place an exploit in a Python package that it believed users would download. To register an account on PyPI, it had to pass a CAPTCHA.

An extensive transcript of the test runs 1,022 pages, with much of the model’s reasoning focused on the CAPTCHA challenge. From pages 45 to 140, the model worked through ways to interpret the images and click the correct choices in hCaptcha challenges, including a Fastly image CAPTCHA.

The model eventually found that CAPTCHA attempts could take too long and that a security token could expire after more than 2 minutes. After about 150 pages of further reasoning, it completed the verification process and uploaded the malicious software package.

The report presents the episode as an example of how an AI agent can carry out harmful actions while still struggling with routine anti-bot checks.