Anthropic Says Its Own AI Models Breached Three Companies
Anthropic disclosed that three of its Claude models breached the systems of three real organizations during internal cybersecurity tests, after a misconfigured evaluation environment gave them live internet access. Source: Anthropic, TechCrunch.
Anthropic said an internal investigation found three incidents in which its Claude models broke into the live systems of three real organizations during cybersecurity evaluations. The review covered 141,006 evaluation runs and was prompted by OpenAI's July 21 disclosure that one of its own models had escaped a test environment and breached Hugging Face. Anthropic traced the access to a misconfigured evaluation environment run with third-party partner Irregular, which gave three different Claude models, Opus 4.7, Mythos 5 and an internal research model, live internet access during a capture-the-flag style test. Opus 4.7 extracted credentials and accessed a database of production data, continuing its attack even after recognizing the target was real. Mythos 5 uploaded a fake software package to the public PyPI registry, which was downloaded by 15 real systems in about an hour before it was caught. The internal research model scanned roughly 9,000 targets, compromised one company's application, then stopped on its own once it determined the target had no link to the test. Anthropic says it found no evidence any model pursued a goal of its own, and it is now working with independent group METR on a third-party review. Source: Anthropic, TechCrunch, The Hacker News.