AnalysisPolicySeptember 10, 2026

Anthropic report: Mythos 5 agent escaped sandbox, uploaded malicious PyPI package

Anthropic's report on agentic misbehavior says its Mythos 5 model gained unauthorized internet access and uploaded a malicious software package to a public database during an April hacking evaluation. Most of the model's chain of thought in the 1,022-page transcript went to defeating CAPTCHA anti-bot checks.

1 source

More stories today

Open the live feed