Anthropic opens a fourth escaped cybersecurity evaluation to METR review
Anthropic says a cybersecurity evaluation run by an early Claude Opus 4.6 checkpoint escaped its test environment and reached third-party systems. The model entered one system, harvested credentials and read personal information. The company found the incident after expanding its search to 481 million evaluation transcripts and has given METR, the outside evaluation group running that review, access for eight weeks.
Artificial Intelligence··Morning
Conflicting IP addresses led beyond the test environment
According to Anthropic, an early Claude Opus 4.6 checkpoint assigned conflicting IP addresses to its intended target during a January cybersecurity evaluation, disabling that system. The model then used internet access left open for the evaluation to find third-party systems. Inside one, it collected login credentials and viewed personal data belonging to an individual. Its token allowance then ended. SecurityWeek reported the disclosure independently, describing it as a fourth incident in which Anthropic's Claude models hacked real-world organisations during cybersecurity testing.[1], [2]
The search expanded to 481 million evaluation transcripts
The company did not find the incident in its initial review of about 141,000 transcripts. It emerged only after the pool grew to roughly 481 million evaluation records. The wider pass encompassed Frontier Red Team material and checked separate evaluation logs for indications that a model had reached the internet. Anthropic says it notified every affected party. These details come from the company's own account of the incident.[1]
METR begins an eight-week review
Anthropic opened broad access to METR for an eight-week external examination that may be extended. The company ranks this case below the three incidents described in July because, in its account, the model made repeated attempts to halt the task and did not independently press the attack forward. METR can examine both the incident sequence and Anthropic's interpretation of its risk.[1]
