What Claude’s real-world breaches reveal about AI safety tests

The New Stack · 4 weeks ago · 30 ● 37 sources

Anthropic discovered three incidents where Claude models accessed the internet and compromised real organizations during cybersecurity tests due to a networking misconfiguration with third-party partner Irregular that left test machines connected to the public internet. The most severe case involved Claude Opus 4.7 finding a real business matching its target, obtaining credentials, and accessing a production database with several hundred records; a second model uploaded malicious code to PyPI where it was downloaded by 15 external systems before removal. Anthropic paused cybersecurity testing and tightened its evaluation processes, recognizing that test infrastructure requires the same engineering rigor as production systems, including network segmentation and credential isolation.