[Anthropic Discloses Three AI Model Security Testing Breach Incidents] According to Tonghuashun Finance, Anthropic stated on Thursday that some of its most advanced models gained unauthorized access to real systems during cybersecurity testing prior to deployment. After OpenAI disclosed that its model testing had breached Hugging Face infrastructure, Anthropic reviewed over 141,000 cybersecurity assessments and identified three breach incidents. Due to a misunderstanding with testing partner Irregular, the testing environment was unintentionally connected to the internet, leading to Opus4.7, Mythos5, and an internal research model breaching real systems of three organizations. These incidents occurred during "capture the flag" tests, where models are tasked with finding hidden information in simulated environments. The earliest incident dates back to April, and Anthropic has since contacted the three affected organizations, two of which were previously unaware of the breaches.