SafetyPage

Lab system cards and frameworks · System card · OpenAI

GPT-5.6 — August Updates

6 Aug 2026

Read at deploymentsafety.openai.com ↗Read at sourceAll rights reserved

Outline · 187 paragraphs

  1. GPT-5.6 — August Updates¶2
  2. Change log¶3
  3. 1. Introduction¶5
  4. 2. Model Data and Training¶10
  5. 3. Safety¶16
  6. 3.1 Disallowed Content¶17
  7. 3.1.1 Evaluations with Challenging Prompts¶18
  8. Table 1: Production Benchmarks with Challenging Prompts (higher is better)¶23
  9. 3.2 Disallowed Content for Users Under-18¶26
  10. Table 2¶35
  11. 3.3 Vision¶36
  12. Table 3: Image input evaluations, with metric not_unsafe (higher is better)¶38
  13. 3.4 Dynamic Mental Health Benchmarks with Adversarial User Simulations¶40
  14. Table 4: Dynamic Benchmarks with Adversarial User Simulations¶45
  15. 4. Robustness¶46
  16. 4.1 Jailbreaks¶47
  17. Figure 1¶53
  18. 4.2 Prompt injection¶56
  19. Table 5: Prompt injection evaluations (higher is better)¶58
  20. 5. Health¶60
  21. 5.1 HealthBench¶61
  22. Table 6: Reported as length-adjusted score (unadjusted, mean response length in characters)¶65
  23. 6. Hallucinations¶68
  24. Figure 2¶74
  25. 7. Preparedness¶76
  26. 7.1 Capabilities Assessment¶81
  27. 7.1.1 Biological and Chemical¶83
  28. Table 7: Overview of Biological and Chemical evaluations¶88
  29. 7.1.1.1 Multimodal Troubleshooting Virology¶90
  30. Figure 3¶94
  31. 7.1.1.2 ProtocolQA Open-Ended¶97
  32. Figure 4¶100
  33. 7.1.1.3 Tacit Knowledge and Troubleshooting¶103
  34. Figure 5¶107
  35. 7.1.1.4 TroubleshootingBench¶110
  36. Figure 6¶114
  37. 7.1.1.5 Hard-negative protein binding prediction¶117
  38. Table 8: Hard-negative protein binding prediction¶123
  39. 7.1.1.6 DNA sequence design for transcription factor binding¶124
  40. Table 9: DNA sequence design for Transcription Factor binding¶129
  41. 7.1.1.7 AAV Capsid Packaging Prediction¶130
  42. 7.1.2 Cybersecurity Capabilities¶132
  43. Table 10: Vulnerability Identification and Exploitation Capabilities evaluations¶136
  44. 7.1.2.1 Cyber Capability Evaluations (Threshold: High)¶138
  45. 7.1.2.1.1 Capture the Flag (CTF) Challenges¶139
  46. Figure 7¶150
  47. 7.1.2.2 CVE-Bench¶153
  48. Figure 8¶156
  49. 7.1.2.3 Cyber range¶159
  50. Table 11¶162
  51. 7.1.3 AI Self-Improvement¶167
  52. 7.2 Safeguards¶169
  53. 7.2.1 Model Safety Training and Evaluation¶172
  54. 7.2.1.1 Biological and Chemical Safety Training and Evaluation¶174
  55. Table 12¶177
  56. 7.2.1.2 Cybersecurity Safety Training and Evaluation¶179
  57. Table 13¶181
  58. 8. References¶182