SafetyPage

Lab system cards and frameworks · System card · OpenAI

GPT-5.5 Instant System Card

5 May 2026

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

Outline · 168 paragraphs

  1. GPT-5.5 Instant System Card¶2
  2. 1. Introduction¶3
  3. 2. Model Data and Training¶6
  4. 3. Safety¶12
  5. 3.1 Disallowed Content¶13
  6. 3.1.1 Evaluations with Challenging Prompts¶14
  7. Table 1: Production Benchmarks with Challenging Prompts (higher is better)¶19
  8. 3.2 Vision¶25
  9. Table 2: Image input evaluations, with metric not_unsafe (higher is better)¶27
  10. 3.3 Dynamic Mental Health Benchmarks with Adversarial User Simulations¶29
  11. Table 3¶30
  12. 4. Robustness¶35
  13. 4.1 Jailbreaks¶36
  14. Figure 1¶41
  15. 4.2 Prompt injection¶43
  16. Table 4: Prompt injection evaluations (higher is better)¶45
  17. 5. Health¶47
  18. 5.1 HealthBench¶48
  19. Table 5: Reported as length-adjusted score (unadjusted, mean response length in characters)¶53
  20. 6. Hallucinations¶56
  21. Figure 2¶62
  22. 7. Inclusivity¶64
  23. 7.1 Bias¶65
  24. Table 6: First-person fairness evaluation (lower is better)¶69
  25. 8. Preparedness¶71
  26. 8.1 Capabilities Assessment¶76
  27. 8.1.1 Biological and Chemical¶78
  28. Table 7: Overview of Biological and Chemical evaluations¶81
  29. 8.1.1.1 Multimodal Troubleshooting Virology¶83
  30. Figure 3¶85
  31. 8.1.1.2 ProtocolQA Open-Ended¶88
  32. Figure 4¶90
  33. 8.1.1.3 Tacit Knowledge and Troubleshooting¶93
  34. Figure 5¶97
  35. 8.1.1.4 TroubleshootingBench¶100
  36. Figure 6¶104
  37. 8.1.2 Cybersecurity¶107
  38. Table 8: Vulnerability Identification and Exploitation Capabilities evaluations¶115
  39. 8.1.2.1 Capture the Flag (CTF) Challenges¶122
  40. Figure 7¶133
  41. 8.1.2.2 CVE-Bench¶136
  42. Figure 8¶139
  43. 8.1.2.3 Cyber range¶142
  44. Table 9¶145
  45. 8.1.3 AI Self-Improvement¶148
  46. 8.2 Safeguards¶150
  47. 8.2.1 Biological Safeguards¶151
  48. Table 10: Below, we share the results of our biological safety evaluations. These evaluations show, for some of the most challenging scenarios that the model can encounter, how often model training alone suffices to prevent a violative response (whether that response is a refusal to provide weaponiz¶153
  49. Table 11: Given the observed regressions on the synthetic evaluations, we subsequently ran our evaluations across both our model and our automated safety monitors to better understand the end-to-end safeguard stack users will encounter in production. We see that the combination of the model itself a¶154
  50. 8.2.2 Cybersecurity Safeguards¶156
  51. Table 12¶160
  52. 9. References¶161