Overview

• Version 3.0 introduces a misalignment risk track that focuses on models that may deceive users, ignore instructions, or resist shutdown or modification.
• Google DeepMind also adds a new Critical Capability Level for harmful manipulation, warning that highly manipulative models could be misused to cause large-scale harm.
• The framework recommends applying automated monitors to inspect a model’s explicit reasoning scratchpads for signs of deception or noncompliance.
• Researchers caution that future systems may reason effectively without producing verifiable chains of thought, limiting the effectiveness of current oversight tools.
• DeepMind notes there is no definitive fix yet and flags broader governance concerns, including the possibility that powerful models could accelerate ML research in untrusted hands.