Google DeepMind Enhances Frontier Safety Framework to Tackle AI Manipulation and Misalignment Risks

Google DeepMind has released the third iteration of its Frontier Safety Framework (FSF), marking a significant step forward in its efforts to responsibly develop advanced AI systems. The updated framework reflects a more comprehensive and evidence-based approach to identifying and mitigating severe risks associated with increasingly powerful AI models. A key addition to this version is the introduction of a new Critical Capability Level (CCL) focused on harmful manipulation. This CCL targets AI models capable of systematically and substantially influencing beliefs and behaviors in high-stakes contexts over time, potentially leading to widespread harm. The update is grounded in DeepMind’s ongoing research into manipulation mechanisms in generative AI, and the company plans to continue investing in understanding and measuring these risks. The framework also evolves its approach to misalignment risks, particularly those involving AI models that could interfere with human control over their operations. While earlier versions included exploratory CCLs related to instrumental reasoning—such as deceptive behavior—this update introduces more defined protocols for managing models that could accelerate AI development to destabilizing levels. It addresses both misuse risks and the dangers of undirected action by highly capable models, especially as they become integrated into AI research and deployment pipelines. To manage these risks, DeepMind now conducts safety case reviews before any external launch when critical capability thresholds are reached. These reviews involve detailed analyses to demonstrate that risks have been reduced to manageable levels. The company is also expanding this rigorous evaluation process to include large-scale internal deployments of advanced models. DeepMind has sharpened its CCL definitions to better target the most severe threats, ensuring that governance and mitigation efforts are proportionate to the risk level. Safety and security measures are now embedded throughout the model development lifecycle, applied proactively before critical thresholds are crossed. The updated framework also provides greater clarity on the risk assessment process. Building on early-warning evaluations, it now includes systematic risk identification, in-depth capability analysis, and explicit judgments on risk acceptability, enabling more holistic and transparent decision-making. This latest version underscores DeepMind’s ongoing commitment to advancing frontier AI safely. By expanding risk domains and refining assessment methods, the company aims to ensure that transformative AI benefits humanity while minimizing potential harms. The framework will continue to evolve based on new research, real-world implementation experience, and input from experts across industry, academia, and government. DeepMind emphasizes that achieving beneficial artificial general intelligence (AGI) requires not only technical innovation but also strong, adaptive safety frameworks. The updated Frontier Safety Framework is intended as a contribution to the global effort to build safe and trustworthy AI.