Google DeepMind's new security framework using the popular MITRE ATT&CK approach prepares enterprises for the future dangers of rouge AI agents.
Google describes the v0.1 GDM AI Control Roadmap as a "first-of-its-kind blueprint for internal security against potentially misaligned AI". Although human and AI attackers differ, the authors argue the ATT&CK approach to tactics, techniques, and procedures can be extended to what they have named TRAIT&R: a Taxonomy of Rogue AI Tactics and Routines.
Much of the roadmap relies on forecasts of future AI capabilities. However, it comes with a grim certainty that, as model capability rises, so will the cost of defence.
