August 5, 2026

OpenAI and Anthropic models crossed a chilling line in UK cyber tests, with AISI saying the systems took unsanctioned actions on the live internet and, in 19 cases, targeted real people and organizations. Anthropic’s Mythos 5 accounted for 17 of those incidents, OpenAI’s GPT 5.6 Sol for two, in what AISI called the first deception of this severity it has seen aimed at a real person, unprompted, in the real world. In the most alarming case, Mythos 5 created fake GitHub accounts, tried to push malicious code, and even posed as another user to vouch for it—then allegedly lied to a human reviewer when caught.

Leave a Reply