OpenAI published a framework document on 26 October 2023 outlining its approach to evaluating catastrophic risks from frontier AI systems [source]. The document describes a four-tier scorecard for assessing model capabilities in cybersecurity, chemical/biological/radiological/nuclear (CBRN) threats, persuasion, and model autonomy.

The framework assigns risk levels from "low" to "critical". OpenAI states it will only deploy models scoring below "high" risk, and that models reaching "critical" — defined as capabilities exceeding those of non-state actors — would trigger security measures equivalent to handling restricted national security information.

The document acknowledges current limitations in risk assessment methodology. OpenAI notes that evaluations may produce false negatives, that model capabilities can emerge unpredictably, and that third-party access to models complicates containment. The framework does not specify quantitative thresholds for each tier, relying instead on qualitative descriptions of threat scenarios.

OpenAI commits to quarterly updates on the highest risk level observed across its model portfolio, with the first assessment rating GPT-4 as "low risk" across all categories. The company states it will revise the framework as understanding of frontier risks evolves.

The publication follows broader industry discussion about voluntary commitments on AI safety. OpenAI indicates the scorecard will inform decisions about model deployment, security protocols, and whether to proceed with training runs. The framework does not address risks outside the four specified domains, such as misinformation at scale, economic disruption, or algorithmic bias in high-stakes applications.