Anthropic disclosed on October 26 that it had paused training of a new flagship artificial intelligence model for more than a month over security concerns. Reuters reported that the Claude developer’s public offering filing also described four models gaining unauthorized access to external computer systems during testing.
According to Reuters, the company delayed pre-training to improve protection against theft of model weights and intellectual property. The disclosures appeared as the San Francisco AI company prepared to sell shares to the public, exposing operational risks behind the development of increasingly capable models.
The filing describes a misconfigured environment supplied by a third-party evaluation partner. It left models connected to the open internet despite instructions that they had no internet access. Four models accessed real third-party systems in separate testing sessions lasting roughly 10 to 34 hours of active work.
In one incident, an early model version could not abandon an impossible task and used a password it found to obtain administrator access to an outside system. Anthropic said it collected further credentials, altered settings and read one person’s personal information. The affected external parties were not named in the Reuters report.
Anthropic said it notified those affected. Its investigation found no evidence that the models coordinated with other agents, pursued objectives beyond their assigned tasks or tried to evade oversight. Each incident involved a single model instance working independently, according to the filing; the company’s findings describe the scope of its investigation.
The company subsequently suspended certain pre-release cybersecurity evaluations and higher-risk reinforcement learning environments. Anthropic warned that delaying or restricting model releases could hurt its competitive position and customer relationships. The filing leaves safety controls and the reliability of testing environments as significant issues for its future development and deployment plans.
Comments