Google DeepMind today is releasing the third iteration of its Frontier Safety Framework, adding a Critical Capability Level for harmful manipulation and expanding its risk assessment to cover misalignment scenarios where models might interfere with operators’ ability to direct, modify or shut down their operations.

The harmful manipulation CCL targets AI systems capable of “systematically and substantially” altering beliefs or behaviors in high-stakes contexts, potentially causing harm at severe scale. The shutdown resistance category addresses risks from models that could interfere with operators’ ability to direct, modify or shut down their operations. DeepMind says it will require safety case reviews before external launches when relevant CCLs are reached, and says large-scale internal deployments can also require review for advanced machine learning research and development CCLs.

The update also refines CCL definitions to trigger governance mechanisms only at the highest threat levels, and describes a more detailed risk assessment process from identification through mitigation. In a blog post, the lab emphasizes that mitigations should apply proactively before systems cross dangerous thresholds.

The third iteration comes 16 months after the framework’s initial release in May 2024. DeepMind says it will continue evolving the framework based on new research, stakeholder input, and implementation lessons.