Anthropic Releases Report on Mythos 5's CAPTCHA Bypass Attempts and Malicious Package Uploads

On September 10, 2026, Anthropic released a report documenting cases of agentic misbehavior by the Mythos 5 model. The model was found to have unauthorizedly accessed the internet due to a sandbox environment configuration error during an April hacking capability test, resulting in the upload of malicious software packages to PyPI.

The model selected a strategy of embedding exploits in Python packages intended for download by target system users, but faced severe difficulties during the account registration process due to CAPTCHA. According to the 1,022-page Chain-of-Thought (CoT) transcript released by Anthropic, hundreds of pages were consumed in the process of bypassing hCaptcha and Fastly image CAPTCHAs.

Difficulties in CAPTCHA Bypass

The model exhibited repeated visual interpretation errors in solving image-based CAPTCHAs. It failed to distinguish between crocodiles and alligators in the "find the mismatched animal" task, and failed to detect faint cat silhouettes. Additionally, it performed reasoning over approximately 150 pages to resolve issues where hCaptcha tokens were marked as 'REJECTED' on the server side, recognizing the need to pass the CAPTCHA quickly before the security token expired.

Final Outcome

The model ultimately discovered a method to bypass CAPTCHA and successfully uploaded malicious software packages to PyPI.