SafetyPage

Lab system cards and frameworks · System card · OpenAI

GPT-5.4 Thinking System Card

5 Mar 2026

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

Outline · 300 paragraphs

  1. GPT-5.4 Thinking System Card¶2
  2. 1. Introduction¶3
  3. 2. Model Data and Training¶6
  4. 3. Baseline Model Safety Evaluations¶11
  5. 3.1 Disallowed Content Evaluations with Challenging Prompts¶12
  6. Table 1: Production Benchmarks with Challenging Prompts (higher is better)¶16
  7. Table 2: Dynamic Benchmarks with Adversarial User Simulations¶18
  8. 3.2 Production Benchmarks with Representative Prompts¶22
  9. Table 3: Disallowed Content Evaluations with Representative Prompts (higher is better)¶28
  10. Figure 1¶30
  11. 3.3 Jailbreaks¶33
  12. Figure 2¶37
  13. 3.4 Prompt injection¶39
  14. Table 4: Prompt injection evaluations¶41
  15. 3.5 Vision¶43
  16. Table 5: Image input evaluations, with metric not_unsafe (higher is better)¶45
  17. 3.6 Health¶47
  18. Table 6: Health evaluations¶50
  19. 3.7 Avoid Accidental Data-Destructive Actions¶54
  20. Table 7: Destructive action avoidance¶56
  21. Table 8¶58
  22. 3.8 User Confirmations During Computer Use¶59
  23. Table 9¶66
  24. 3.9 Bias¶67
  25. Table 10: First-person fairness evaluation (lower is better)¶71
  26. 4. Chain of Thought Evaluations¶72
  27. 4.1 CoT Monitorability¶74
  28. Figure 3¶81
  29. Figure 4¶83
  30. Figure 5¶85
  31. Figure 6¶87
  32. 4.2 CoT Controllability¶89
  33. Figure 7¶93
  34. 5. Preparedness Framework¶95
  35. 5.1 Capabilities Assessment¶99
  36. 5.1.1 Biological and Chemical¶101
  37. Table 11: Overview of Biological and Chemical evaluations¶104
  38. 5.1.1.1 Multi-select Multimodal Troubleshooting Virology¶106
  39. Figure 8¶108
  40. 5.1.1.2 ProtocolQA Open-Ended¶111
  41. Figure 9¶113
  42. 5.1.1.3 Tacit Knowledge and Troubleshooting¶116
  43. Figure 10¶120
  44. 5.1.1.4 TroubleshootingBench¶123
  45. Figure 11¶127
  46. 5.1.2 Cybersecurity¶130
  47. Table 12: Vulnerability Identification and Exploitation Capabilities evaluations¶139
  48. 5.1.2.1 Capture the Flag (CTF) Challenges¶146
  49. Figure 12¶157
  50. 5.1.2.2 CVE-Bench¶159
  51. Figure 13¶162
  52. 5.1.2.3 Cyber range¶165
  53. Table 13: Cyber Range¶168
  54. Figure 14¶176
  55. 5.1.2.4 External Evaluations for Cyber Capabilities¶179
  56. 5.1.3 AI Self-Improvement¶188
  57. Table 14: Overview of AI Self-Improvement evaluations¶190
  58. 5.1.3.1 Monorepo-Bench¶192
  59. Figure 15¶198
  60. 5.1.3.2 MLE-Bench¶201
  61. Figure 16¶203
  62. 5.1.3.3 OPQA¶206
  63. Figure 17¶208
  64. 5.2 Research Category Update: Sandbagging¶211
  65. 5.3 Cyber Safeguards¶218
  66. 5.3.1 Threat Model and Scenarios¶226
  67. 5.3.2 Cyber Threat Taxonomy¶228
  68. 5.3.3 Model Safety Training¶234
  69. Table 15¶237
  70. 5.3.4 Conversation monitor¶240
  71. 5.3.5 Actor Level Enforcement¶244
  72. 5.3.6 Trust-based access¶250
  73. 5.3.7 Security Controls¶252
  74. 5.3.8 Misalignment risks and internal deployment¶254
  75. 6. Appendix: GPT-5.4 mini¶256
  76. 6.1 Disallowed Content¶258
  77. Table 16: Production Benchmarks with Challenging Prompts (higher is better)¶259
  78. Table 17: Dynamic Benchmarks with Adversarial User Simulations (higher is better)¶260
  79. 6.2 GPT-5.4 mini CoT controllability results¶261
  80. Figure 18¶263
  81. 6.3 Preparedness Framework¶265
  82. 6.3.1 Biological and Chemical¶267
  83. Table 18¶276
  84. 6.3.2 Cybersecurity¶279
  85. Table 19¶280
  86. 6.3.3 AI Self Improvement¶282
  87. Table 20¶283
  88. 7. References¶285