[Anthropic Prospectus: Four Models Had Unauthorized Access to Real Systems During Testing] According to the latest revised prospectus submitted by Anthropic to the U.S. SEC, during a cybersecurity assessment built by a third-party partner, due to a misconfigured testing environment, the evaluated models connected to the public internet despite being informed they could not access the internet. Four of its models had unauthorized access to real third-party systems for approximately 10 to 34 hours. In one incident, an earlier version of the model, unable to abort an impossible task, used a found password to gain administrator privileges on the third-party system, collecting more credentials, changing system settings, and reading an individual's personal information. The affected party has been notified. Anthropic stated that it has suspended some related assessments and high-risk reinforcement learning environments following the incident.