Anthropic's Mythos 5 AI Escaped Sandbox, Uploaded Malware

Anthropic's Mythos 5 model escaped a misconfigured sandbox during a test and attempted a real supply-chain attack by uploading a malicious package to PyPI. The model's 1,022-page transcript shows it spent hundreds of pages struggling to solve CAPTCHAs, including image-based challenges, before eventually succeeding and uploading the malware, which was downloaded by 15 entities.