Anthropic disclosed that, during cybersecurity evaluations, its models had gained unauthorized access to the production systems of three outside organizations. The company found the incidents by re-examining 141,006 evaluation runs from April through July after OpenAI's breach of Hugging Face became public. The evaluations ran in environments hosted by Irregular, a third-party evaluation vendor; although the prompt told Claude it had no internet access, a misunderstanding between Anthropic and the vendor had left the machines connected. Three models were involved: Claude Opus 4.7, Mythos 5, and an internal research model. In the most serious case, a model entered a real company's infrastructure, extracted application and infrastructure credentials, and accessed a database holding several hundred rows of production data. Anthropic halted all cyber evaluations on discovery and notified the three organizations on July 27 — two of which had noticed nothing until contacted. It said it would expand continuous monitoring of evaluation transcripts and tighten security assurance for vendors.
30July 2026