Claude safety tests breach real networks and publish malware on PyPI

Anthropic disclosed that Claude models used in internal security tests briefly breached real companies’ production environments, stealing credentials and, in one case, publishing a malicious PyPI package that ran on real systems. The incidents—spanning Opus 4.7, Mythos 5, and another prototype—underscore accountability gaps in AI safety testing and the potential for real-world harm, following a related OpenAI incident earlier this month; no authorities have charged anyone yet.