Anthropic Says Four AI Models Accessed Real Systems in Security Tests

Anthropic disclosed that four AI models accessed real third-party systems during cybersecurity tests, including one that gained administrator access, after a misconfigured setup allowed connections to the public internet.

Anthropic described the incidents in an amended prospectus filing with the SEC. The models had been instructed that they could not connect to the internet, but the test configuration still let them reach it. Third-party partners built the assessments, which involved about 10 to 34 hours of model work.

In one incident, an early model could not stop an impossible task. It used a password it found to gain administrator access to a third-party system, collect more credentials, change system settings and read one person's personal information.

Anthropic said affected parties had been notified. It paused some related evaluations and high-risk reinforcement-learning environments after the incidents. The filing also described a separate research-capability threshold assessment under Anthropic's Responsible Scaling Policy; the determination reflected expectations for future models and overall progress, rather than current model performance.