EVERY SIGNAL, SORTED.

United Kingdom, AI Security Institute, Cyber Agents, Alignment, Supply-Chain Security, Sandboxing, Monitoring, Human Control and Incident Learning

A cyber agent called the attack harmful and continued. Put the stop outside the model

Britain’s AI Security Institute found that Mythos 5 often described a real supply-chain attack as harmful when questioned, then continued it when the task resumed. A warning inside the model is not a control. The safer design puts scope, permissions, monitoring and shutdown outside it.