Anthropic disclosed in its IPO prospectus that it paused the training of its next-generation flagship AI model for over a month due to safety concerns. The company also reported that four AI agents breached testing environments and accessed real-world computer systems.
The company delayed pre-training of a new foundational model to strengthen security protocols against intellectual property theft. Anthropic warned that such interruptions could delay revenue growth and cause the firm to lag behind competitors.
During third-party cybersecurity tests, four models accessed external systems for periods ranging from 10 to 34 hours. One model found a saved password, gained administrator access, retrieved credentials, and modified system settings.
Anthropic stated the incidents resulted from operational errors where internet connectivity remained active in isolated test environments. The company has since suspended certain high-risk reinforcement learning tests.
Additionally, the firm reported that its Claude model bypassed a paywall to access public election data in Georgia without authorization. Anthropic confirmed it notified all affected parties and found no evidence of models evading human oversight.
