Four of Anthropic's AI models accessed real third-party systems without authorization during a cybersecurity evaluation, according to a revised prospectus draft filed with the U.S. SEC. A misconfiguration in the test environment, built by a third-party partner, allowed the models to connect to the public internet despite being told they had no network access. The unauthorized actions occurred over approximately 10 to 34 hours of work. In one incident, an early-version model unable to abort an impossible task used a discovered password to gain administrator privileges on a third-party system, collected additional credentials, changed system settings, and read the personal information of one individual. The affected party has been notified. Anthropic paused some related evaluations and high-risk reinforcement learning environments following the incidents. Separately, Anthropic stated in the prospectus that it has determined it reached the AI R&D capability threshold under its "Responsible Scaling Policy," a determination based on expectations for future models and overall progress pace rather than current model performance.