SafetyPage

Lab system cards and frameworks · System card · OpenAI

GPT-5.5 System Card

23 Apr 2026

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

Outline · 343 paragraphs

  1. GPT-5.5 System Card¶2
  2. Change log¶3
  3. 1. Introduction¶5
  4. 2. Model Data and Training¶10
  5. 3. Safety¶17
  6. 3.1 Disallowed Content¶18
  7. 3.1.1 Evaluations with Challenging Prompts¶19
  8. Table 1: Production Benchmarks with Challenging Prompts (higher is better)¶23
  9. 3.1.2 Evaluations with Representative Prompts¶27
  10. Figure 1¶34
  11. 3.2 Vision¶36
  12. Table 2: Image input evaluations, with metric not_unsafe (higher is better)¶38
  13. 3.3 Avoiding Accidental Data-Destructive Actions¶40
  14. Table 3¶42
  15. Table 4¶43
  16. 3.4 User Confirmations During Computer Use¶45
  17. Table 5¶51
  18. 4. Robustness Evaluations¶52
  19. 4.1 Jailbreaks¶53
  20. Figure 2¶56
  21. 4.2 Prompt injection¶58
  22. Table 6: Prompt injection evaluations¶60
  23. 5. Health¶61
  24. 5.1 HealthBench¶62
  25. Table 7: Reported as length-adjusted score (unadjusted, mean response length in characters)¶67
  26. 5.2 Dynamic Mental Health Benchmarks with Adversarial User Simulations¶70
  27. Table 8: Dynamic Benchmarks with Adversarial User Simulations¶73
  28. 6. Hallucinations¶74
  29. 6.1 Performance in Cases Flagged by Users¶75
  30. Figure 3¶78
  31. 7. Alignment¶80
  32. 7.1 Evaluations with Prompts Representative of External ChatGPT Usage¶81
  33. Figure 4¶84
  34. 7.2 Evaluating Coding Agents for Misalignment by Resampling Internal Traffic¶86
  35. Figure 5¶89
  36. 7.2.1 Measuring our ability to detect misalignment¶95
  37. Figure 6¶96
  38. 7.3 Chain of Thought Evaluations¶101
  39. 7.3.1 CoT Monitorability¶102
  40. Figure 7¶109
  41. Figure 8¶111
  42. Figure 9¶113
  43. Figure 10¶115
  44. 7.3.2 CoT Controllability¶117
  45. Figure 11¶121
  46. 8. Bias Evaluation¶123
  47. 8.1 First Person Fairness Evaluation¶124
  48. Table 9: First-person fairness evaluation (lower is better)¶128
  49. 9. Preparedness¶129
  50. 9.1 Capabilities Assessment¶133
  51. 9.1.1 Biological and Chemical¶135
  52. Table 10: Overview of Biological and Chemical evaluations¶138
  53. 9.1.1.1 Multimodal Troubleshooting Virology¶140
  54. Figure 12¶142
  55. 9.1.1.2 ProtocolQA Open-Ended¶145
  56. Figure 13¶147
  57. 9.1.1.3 Tacit Knowledge and Troubleshooting¶150
  58. Figure 14¶154
  59. 9.1.1.4 TroubleshootingBench¶157
  60. Figure 15¶161
  61. 9.1.1.5 Biochemistry knowledge improvement over GPT-5.4-thinking¶163
  62. Table 11: Biochemistry knowledge improvement over GPT-5.4-thinking¶165
  63. 9.1.1.6 Hard-negative protein binding prediction¶167
  64. Table 12: Hard negative protein binding prediction¶174
  65. 9.1.1.7 DNA sequence design for transcription factor binding¶175
  66. Table 13: DNA sequence design for Transcription Factor binding¶178
  67. 9.1.1.8 External Evaluation for Bio Capabilities - SecureBio¶179
  68. 9.1.1.9 External Evaluations for Bio Capabilities - US CAISI¶185
  69. 9.1.1.10 Bio Bug Bounty Program¶188
  70. 9.1.2 Cybersecurity¶190
  71. Table 14: Vulnerability Identification and Exploitation Capabilities evaluations¶195
  72. 9.1.2.1 Capture the Flag (CTF) Challenges¶197
  73. Figure 16¶208
  74. 9.1.2.2 CVE-Bench¶210
  75. Figure 17¶213
  76. 9.1.2.3 Cyber range¶216
  77. Table 15¶219
  78. 9.1.2.4 VulnLMP¶227
  79. 9.1.2.5 External Evaluations for Cyber Capabilities - Irregular¶232
  80. 9.1.2.6 External Evaluations for Cyber Capabilities - US CAISI¶237
  81. 9.1.2.7 External Evaluations for Cyber Capabilities - UK AISI¶240
  82. 9.1.3 AI Self-Improvement¶246
  83. Table 16: Overview of AI Self-Improvement evaluations¶248
  84. 9.1.3.1 Monorepo-Bench¶250
  85. Figure 18¶256
  86. 9.1.3.2 MLE-Bench¶259
  87. Figure 19¶261
  88. 9.1.3.3 Internal Research Debugging Evaluation¶264
  89. Figure 20¶266
  90. Figure 21¶270
  91. 9.1.3.4 OPQA¶272
  92. Figure 22¶274
  93. 9.2 Research Category Update: Sandbagging¶277
  94. 9.2.1 External Evaluations for Sandbagging - Apollo Research¶278
  95. 9.3 Safeguards¶285
  96. 9.3.1 Biological and Chemical Safeguards¶286
  97. 9.3.1.1 Bio Safeguards Testing¶288
  98. 9.3.2 Cyber Safeguards¶293
  99. 9.3.2.1 Threat Model and Scenarios¶297
  100. 9.3.2.2 Model Safety Training¶299
  101. Table 17¶302
  102. 9.3.2.3 Conversation monitor¶303
  103. 9.3.2.4 Actor Level Enforcement¶308
  104. 9.3.2.5 Trust-based access¶315
  105. 9.3.2.6 Security Controls¶317
  106. 9.3.2.7 Cyber Safeguard Testing¶319
  107. 9.3.2.7.1 UK AISI Cyber Safeguard Testing:¶320
  108. 9.3.2.7.2 External Red-teaming Campaigns:¶322
  109. 9.3.2.8 Cyber Frontier Risk Council¶325
  110. 9.3.2.9 Misalignment risks and internal deployment¶328
  111. 10. References¶330