PRISM Risk Signal Framework: Hierarchy-Based Red Lines for AI Behavioral Risk

Red lines can be set at a deeper level than individual cases: Rather than banning specific prompts or outputs one-by-one, the PRISM framework identifies problematic patterns in how AI systems prioritize values, evaluate evidence, and trust information sources—meaning a single structural safeguard can prevent unlimited harmful variations of the same underlying reasoning flaw.

The framework is anticipatory rather than reactive: By detecting dangerous reasoning structures before they produce harmful outputs, PRISM can catch risks earlier in the development process. The paper tested this across approximately 397,000 forced-choice responses from 7 AI models, successfully identifying models with "structurally extreme profiles" before those flaws manifested as specific harms.

The approach is measurable and objective: Unlike subjective judgments about what constitutes harmful AI behavior, PRISM grounds safety assessments in empirical data using a "dual-threshold principle" that combines absolute ranking and performance gaps—making it possible to compare AI systems' reasoning integrity on consistent, quantifiable terms rather than case-by-case judgment calls.

Different AI models show different risk patterns: The framework successfully distinguished between three categories of models: those with extreme structural problems, those with context-dependent risks (safe in some situations but not others), and those with balanced reasoning hierarchies—suggesting that one-size-fits-all AI safety approaches may miss important differences between systems.

The taxonomy identifies 27 specific behavioral risk signals: These signals span three levels of how AI systems make decisions (value prioritization, evidence weighting, and source trust), providing a structured way to diagnose why an AI system might behave dangerously, not just that it does.