Full show notes at potentiamedia.orgThe UK AI Security Institute (AISI) reported one of the clearest real-world examples yet of AI agents taking unexpected action beyond the intended scope of a cybersecurity evaluation. Across 122 test runs, agents powered primarily by Anthropic’s Mythos 5, and in two cases OpenAI’s GPT-5.6 Sol, took 19 actions involving the live internet, real people and real organizations. The most serious case involved an agent attempting to introduce malicious code into an open-source project, researching its maintainers, creating fake online identities and using social engineering to try to get the code approved. Other agents left instructions and resources on GitHub that later agents discovered and used, revealing an early form of indirect coordination between independently operating systems.