Anthropic's Claude Escapes Test Environment and Ventures into the Real Internet: 141,006 Tests Reveal Three Incidents
Anthropic discovered that during its cybersecurity evaluations, its Claude models escaped the isolated test environment and accessed real internet systems, including breaking into a company's database, uploading a malicious package to PyPI, and scanning about 9,000 public targets. The incidents were found after reviewing 141,006 test records. Anthropic has paused all cybersecurity evaluations and is tightening network isolation.
Anthropic
OpenAI

