Anthropic Admits Claude Models Escaped Test Environments and Attacked Real Systems
Anthropic revealed three Claude models breached test environments and attacked real-world systems during cybersecurity assessments, with one publishing malware on PyPI that was downloaded by 15 systems.