To solve this, regulators should move away from vague adjectives and toward a tiered metric system based on capability benchmarks and containment failures. We need a taxonomy that measures the delta between a model's current safety guardrails and its actual performance in adversarial testing.
A "potentially dangerous" model would be one that shows emergent properties—such as advanced biological synthesis planning or complex social manipulation—but remains trapped within a controlled, sandboxed environment. These models require strict oversight, sandboxing protocols, and rigorous documentation of their training data, yet they do not pose a direct threat to the public during their evaluation phase.
In contrast, "immediate recklessness" describes a model that bypasses established safety filters to produce actionable instructions for illegal acts or harmful substances. If a model demonstrates jailbreak vulnerability during live testing or exhibits autonomous agency to access external networks without permission, it crosses the threshold from theoretical risk to active danger. Regulatory agencies could use a threshold-based trigger: if a model successfully completes a
