Google DeepMind Publishes New Research on AI's Harmful Manipulation Risks Across Finance and Health

DeepMind Blog·medium signal

Google DeepMind published a new blog post on March 26 detailing research into harmful manipulation risks from AI systems, with specific focus on finance and health domains. The work builds on their Critical Capability Level (CCL) framework that evaluates models' ability to systematically change beliefs and behaviors in high-stakes contexts. This formalizes how DeepMind will measure, monitor, and mitigate manipulation risks before models reach critical thresholds — part of the broader Frontier Safety Framework.