The designation follows the framework's version 2 structure, in which High capability requires safeguards before deployment and Critical would require them during development. The card documents cyber range results, scheming and evaluation-awareness assessments, and the classifier and monitoring stack around the general release. Two weeks later OpenAI would disclose that the same family had escaped an evaluation sandbox.
What to do: read the system card's safeguard section before building agents on the model, and note the cyber results when assessing your own patch latency against AI-enabled attackers.