SafetyPage

Lab system cards and frameworks · System card · OpenAI

GPT-5.1-Codex-Max System Card

18 Nov 2025

Read at openai.com ↗Read at sourceAll rights reserved

Outline · 27 pages

  1. Introductionp. 4
  2. Baseline model safety evaluationsp. 4
  3. Disallowed content evaluationsp. 4
  4. Jailbreaksp. 5
  5. Visionp. 5
  6. Product-specific risk mitigationsp. 6
  7. Agent sandboxp. 6
  8. Network accessp. 7
  9. Model-specific risk mitigationsp. 7
  10. Harmful tasksp. 7
  11. Risk descriptionp. 7
  12. Mitigationp. 7
  13. Safety trainingp. 7
  14. Prompt injectionp. 8
  15. Risk descriptionp. 8
  16. Mitigationp. 8
  17. Safety trainingp. 8
  18. Avoid data-destructive actionsp. 9
  19. Risk descriptionp. 9
  20. Mitigationp. 9
  21. Safety trainingp. 9
  22. Preparednessp. 10
  23. Capabilities assessmentp. 10
  24. Biological and chemicalp. 10
  25. Long-form biological risk questionsp. 10
  26. Multimodal troubleshooting virologyp. 11
  27. ProtocolQA open-endedp. 11
  28. Tacit knowledge and troubleshootingp. 12
  29. TroubleshootingBenchp. 12
  30. Cybersecurityp. 13
  31. Capture-the-flag (professional)p. 15
  32. CVE-Benchp. 16
  33. Cyber Rangep. 17
  34. External evaluations by Irregularp. 19
  35. Preparing for high cyber capabilityp. 19
  36. AI self-improvementp. 20
  37. SWE-Lancerp. 20
  38. Paperbench-10 (n=10)p. 21
  39. MLE-bench-30 (n=30)p. 22
  40. OpenAI PRsp. 23
  41. OpenAI-Proof Q&Ap. 24
  42. External evaluations by METRp. 25
  43. Research category update: sandbaggingp. 27
  44. External evaluations by Apollo Researchp. 27