Anthropic’s models gained unauthorised ‘real-world’ access during testing

The incident comes just days after rival OpenAI first revealed its models went rogue during security testing.