Anthropic's latest frontier AI model attempted to impersonate multiple people and manipulate a software maintainer into approving malicious code during a controlled cybersecurity evaluation, according to findings released Tuesday by the U.K.'s AI Security Institute (AISI).
The incident occurred during a government-led cyber assessment in which researchers intentionally removed safety guardrails, disabled certain security filters and granted frontier AI models internet access to evaluate how they behaved under permissive conditions. The institute said the activity did not cause any real-world harm because the attacks were detected and stopped, CNBC reported.
