SafetyPage

Lab system cards and frameworks · System card · OpenAI

GPT-5 System Card

13 Aug 2025

Read at openai.com ↗Read at sourceAll rights reserved

Outline · 60 pages

  1. Introductionp. 5
  2. Model Data and Trainingp. 6
  3. Observed Safety Challenges and Evaluationsp. 6
  4. From Hard Refusals to Safe-Completionsp. 6
  5. Disallowed Contentp. 7
  6. Sycophancyp. 8
  7. Looking aheadp. 9
  8. Jailbreaksp. 9
  9. Instruction Hierarchyp. 10
  10. Prompt Injectionsp. 11
  11. Hallucinationsp. 12
  12. Deceptionp. 14
  13. Monitoring Chain of Thought for Deceptionp. 16
  14. Image Inputp. 17
  15. Healthp. 17
  16. Multilingual Performancep. 19
  17. Fairness and Bias: BBQ Evaluationp. 20
  18. Red Teaming & External Assessmentsp. 21
  19. Expert Red Teaming for Violent Attack Planningp. 21
  20. Expert and Automated Red Teaming for Prompt Injectionsp. 22
  21. Preparedness Frameworkp. 23
  22. Capabilities Assessmentp. 23
  23. Biological and Chemicalp. 24
  24. Long-form Biological Risk Questionsp. 24
  25. Multimodal Troubleshooting Virologyp. 25
  26. ProtocolQA Open-Endedp. 25
  27. Tacit Knowledge and Troubleshootingp. 26
  28. TroubleshootingBenchp. 27
  29. External Evaluations by SecureBiop. 27
  30. Cybersecurityp. 28
  31. Capture the Flag (CTF) Challengesp. 29
  32. Cyber rangep. 30
  33. External Evaluations by Pattern Labsp. 32
  34. AI Self-Improvementp. 36
  35. SWE-bench Verified (N=477)p. 36
  36. OpenAI PRsp. 37
  37. MLE-Benchp. 38
  38. SWE-Lancerp. 39
  39. PaperBenchp. 40
  40. OPQAp. 40
  41. External Evaluations by METRp. 42
  42. Research Category Update: Sandbaggingp. 44
  43. External Evaluations by Apollo Researchp. 44
  44. Safeguards for High Biological and Chemical Riskp. 46
  45. Threat model and biological threat taxonomyp. 46
  46. Safeguard designp. 47
  47. Model trainingp. 47
  48. System-level protectionsp. 48
  49. Account-level enforcementp. 48
  50. API accessp. 48
  51. Trusted Access Programp. 49
  52. Safeguard testingp. 49
  53. Testing model safety trainingp. 49
  54. Testing system-level protectionsp. 50
  55. Expert Red Teaming for Bioweaponizationp. 51
  56. Third party red teamingp. 52
  57. External government red teamingp. 53
  58. Security controlsp. 53
  59. Sufficiency of Risk Mitigation Measuresp. 54
  60. Appendix 1p. 55
  61. Appendix 2: Hallucinationsp. 56