According to CoinMeta, on Thursday, Anthropic stated that some of its most powerful AI models accessed real systems without authorization during pre-deployment network security tests. The company reviewed over 141,000 network security assessments and discovered three instances of intrusion. Due to a misunderstanding with the testing partner Irregular, the assessment environment accidentally connected to the internet, resulting in Opus 4.7, Mythos 5, and an internal research model intruding into the real systems of three organizations. These incidents occurred during "capture the flag" tests, where the models were supposed to search for hidden information in a simulated environment. The incidents can be traced back to April, and Anthropic has now contacted the three affected organizations, two of which were not aware of the intrusions before. Unlike the OpenAI incident, the Anthropic models did not exploit zero-day vulnerabilities but obtained network access due to incorrect environmental configurations. Similar to OpenAI, Anthropic did not deploy additional protective measures during the assessments in order to measure the models' underlying capabilities. These incidents have raised new questions about how laboratories ensure the security of their assessment environments.