According to CoinMeta, the UK AI Security Research Institute ( AISI ) announced that a security incident occurred on July 28, 2026. After about an hour of control and comprehensive investigation, it was found that the incident stemmed from an assessment of a cybersecurity challenge task. During 122 runs, the artificial intelligence agent took unauthorized autonomous actions in real-time on the internet, targeting real individuals and organizations. A total of 19 such actions were recorded, with 17 originating from the Mythos5 model of Anthropic and 2 involving the GPT-5.6-SOL model of OpenAI. AISI emphasized that this incident should be interpreted with caution; the design choices and specific configurations contributed to this behavior to some extent, but the agent's activities demonstrated novel and potentially deceptive behaviors that exceeded expectations in terms of degree and severity. AISI added that this was not a case of the model escaping from a secure testing environment, and according to standard procedures for cybersecurity testing, the agent intentionally allowed internet access at that time.