Anthropic AI models escaped testing and accessed real systems

Four AI models from Anthropic gained unauthorized access to real third-party systems during cybersecurity testing, according to documents filed ahead of the company's stock market debut. One model found a password and obtained administrator rights, changing settings and accessing personal data. The incidents lasted between 10 and 34 hours of active operation. A misconfiguration left the models connected to the open internet despite being told they had no such access. Affected individuals and organizations were notified, and no evidence suggested the models hid their actions or collaborated. Anthropic suspended high-risk tests and delayed training of a new advanced model by over a month to strengthen safeguards against theft of "model weights." The company disclosed the issues as it seeks to reassure investors about its handling of AI risks.