1、GPT-5 System CardOpenAIAugust 7,20251Contents1Introduction42Model Data and Training53Observed Safety Challenges and Evaluations53.1From Hard Refusals to Safe-Completions.53.2Disallowed Content.63.3Sycophancy.73.3.1Looking ahead.83.4Jailbreaks.83.5Instruction Hierarchy.93.6Hallucinations.103.7Decepti

2、on.123.7.1Monitoring Chain of Thought for Deception.143.8Image Input.153.9Health.153.10 Multilingual Performance.173.11 Fairness and Bias:BBQ Evaluation.184Red Teaming&External Assessments194.1Expert Red Teaming for Violent Attack Planning.194.2Expert and Automated Red Teaming for Prompt Injections.

3、205Preparedness Framework215.1Capabilities Assessment.215.1.1Biological and Chemical.225.1.1.1Long-form Biological Risk Questions.225.1.1.2Multimodal Troubleshooting Virology.235.1.1.3ProtocolQA Open-Ended.2315.1.1.4Tacit Knowledge and Troubleshooting.245.1.1.5TroubleshootingBench.255.1.1.6External

4、Evaluations by SecureBio.255.1.2Cybersecurity.265.1.2.1Capture the Flag(CTF)Challenges.275.1.2.2Cyber range.285.1.2.3External Evaluations by Pattern Labs.305.1.2.4SWE-bench Verified.345.1.2.5OpenAI PRs.355.1.2.6MLE-Bench.355.1.2.7SWE-Lancer.375.1.2.8PaperBench.385.1.2.9OPQA.395.1.2.10 External Evalu

5、ations by METR.395.2Research Category Update:Sandbagging.415.2.1External Evaluations by Apollo Research.425.3Safeguards for High Biological and Chemical Risk.435.3.1Threat model and biological threat taxonomy.445.3.2Safeguard design.445.3.2.1Model training.455.3.2.2System-level protections.455.3.2.3

6、Account-level enforcement.465.3.2.4API access.465.3.2.5Trusted Access Program.475.3.3Safeguard testing.475.3.3.1Testing model safety training.475.3.3.2Testing system-level protections.485.3.3.3Expert Red Teaming for Bioweaponization.485.3.3.4Third party red teaming.5025.3.3.5External government red