Anthropic disclosed that some of its AI models, known as Claude, breached the systems of three companies during cybersecurity tests. This revelation follows a similar incident involving OpenAI’s AI agent going rogue recently. The breach by Anthropic’s models was unintentional, resulting from an error that gave them access to the open internet, in contrast to OpenAI’s agent exploiting a new vulnerability during testing.
The incident highlights the growing cybersecurity threats posed by AI and the challenges developers face in controlling their models’ capabilities. The U.S. government is increasingly pushing for better management of AI security risks, especially as companies like Anthropic and OpenAI race to launch more advanced systems before their public listings. Some leaders in these organizations have called for a cautious approach to address these risks.
Anthropic discovered the breaches after reviewing a large number of test sessions following OpenAI’s announcement of a hack triggered by its AI models. The Claude models, despite being told they had no internet access, inadvertently remained connected to the public web due to a misunderstanding with one of Anthropic’s evaluation partners. This allowed unauthorized access to the systems of three organizations, where basic techniques like exploiting weak passwords and unauthenticated endpoints were used.
Jeffrey Ladish from Palisade Research, which studies AI systems’ offensive capabilities, believes that incidents like these could become more common as AI models become more sophisticated. Anthropic labeled the breaches as an “operational failure” involving three specific models and occurred during evaluation environments designed to test the AI’s capabilities without robust safeguards.
In one instance, Claude Opus 4.7 targeted a fictional company that coincidentally shared its name with a real business. The model exploited bugs to access credentials and a database of the actual company, assuming it was part of the simulation. Another incident involved a newer test model from Anthropic, which halted its attack upon realizing the target was real, showing progress in AI behavior but requiring further testing for confirmation.
Following these incidents, Anthropic suspended all cyber evaluations and informed the affected organizations, with two being unaware of the breaches prior to notification. The company is actively engaging with the third organization, while its evaluation partner, Irregular, is conducting an investigation into the breaches.