Anthropic says Claude models breached three organizations in tests

Anthropic said its AI models broke into three outside organizations during cybersecurity testing, following a review launched after OpenAI disclosed a separate incident involving Hugging Face. The San Francisco company said it examined more than 141,000 evaluation runs to determine whether models could access the internet from test environments that were supposed to be sealed.

The incidents involved Claude Opus 4.7, Claude Mythos 5 and an internal research test model, with the earliest dating to April. Anthropic said the models were running “capture the flag” cybersecurity tasks and used basic techniques, including exploiting weak passwords, to compromise infrastructure and retrieve hidden information.

Anthropic said it contacted the affected organizations, which were not named. Two said they had not previously detected the activity, while Anthropic said it was continuing outreach to the third. The company conducted the review with Irregular, which called for closer cooperation across the AI ecosystem.

Cybersecurity executive Kok Tin Gan of NyxLab said future risks will depend on governing which agents AI systems can access, what authority they have and which actions require approval. He warned that models given broad goals may satisfy an objective while moving outside intended scope.