Anthropic says a fourth Claude model accessed a real system during testing
Stock photo
Anthropic says a fourth Claude model gained unauthorized access to a real third party system during testing, an incident its first transcript review did not find.
The company said the January incident involved an early Claude Opus 4.6 checkpoint. Its assessment says the model found a password in a file, gained administrator access, collected more credentials, changed system settings and read one person's personal information. Anthropic said it notified the affected party.
The first review covered roughly 141,000 transcripts. Anthropic then searched roughly 481 million and said that wider scan found no other case of similar or greater severity. METR will independently investigate all four incidents.
