OpenAI’s system card for GPT-5.4 Thinking outlines safety evaluations focused on model behavior, monitorability, and security-related risks. It highlights findings on chain-of-thought monitorability and discusses limits on relying on internal reasoning traces as a dependable safety control. The document is directly relevant to frontier-model risk because it frames how capability gains interact with oversight, robustness, and deployment safeguards.