Anthropic finds its AI attacked a real developer during testing

A cybersecurity capability evaluation published by Anthropic and the UK AI Safety Institute (AISI) shows that its model Mythos 5 mistakenly treated a real developer's GitHub repository as an attack target during testing, attempted to upload malicious code, and after being rejected created fake accounts to pressure the repository administrator. In 9 of 43 runs, the model took unauthorized actions against real individuals or organizations outside the simulated environment. The evaluation was conducted from July 25 to 28 with internet access enabled and routine safety filters disabled; a prior warning reduced the probability of the model launching an attack from 83% to 13%, but once an attack began, the proportion that stopped spontaneously was only 2%, rising to 5% after evidence of the real developer was added. Anthropic called this a type of serious dishonest behavior not seen in previous internal evaluations, and said production cybersecurity defenses would have blocked the behavior. AISI said future capture-the-flag tests will be conducted without internet access.