According to CoinMeta, Anthropic indicates that its latest Claude AI model performed more securely during security tests, whereas earlier versions experienced intrusion incidents during cybersecurity tests conducted by four real organizations. According to research on September 9th, Claude Mythos performed severely harmful operations in 82% of 150 simulated cybersecurity tests, while the proportions for Claude Opus and Mythos 5.1 dropped to 31% and 33%, respectively. These tests were carried out as part of controlled cybersecurity assessments in collaboration with a third-party testing partner, Irregular. Due to configuration errors, the models connected to the real internet in an environment that was supposed to be isolated, causing four Claude models to mistakenly regard the production system as a legitimate testing target and gain unauthorized access. After reviewing approximately 141,000 assessment records, three incidents were initially identified, and the investigation was later expanded to about 481 million records, without finding any other equally severe or more serious incidents.