SafetyPage

Lab system cards and frameworks · System card · OpenAI

Addendum to GPT-6 Astra System Card: GPT-6.1 Terra

26 Oct 2026

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

Outline · 255 paragraphs

  1. Addendum to GPT-6 Astra System Card: GPT-6.1 Terra¶2
  2. 1. Addendum to GPT-6 Astra System Card: GPT-6.1 Terra¶3
  3. 2. Model Data and Training¶7
  4. 3. Model Safety¶13
  5. 3.1 Safe Completions¶15
  6. 3.1.1 Evaluations with Challenging Prompts¶16
  7. Table 1¶20
  8. Figure 1¶22
  9. 3.1.2 Safe Completions for Users Under 18¶24
  10. Table 2¶29
  11. 3.1.3 Agentic Safe Completions¶31
  12. Table 3¶34
  13. 3.2 Vision¶36
  14. Table 4¶39
  15. 4. Robustness¶41
  16. 4.1 Jailbreaks¶42
  17. Figure 2¶46
  18. 4.2 Prompt Injection¶48
  19. Figure 3¶51
  20. 5. Health¶53
  21. 5.1 HealthBench¶54
  22. Table 5¶57
  23. 5.2 Dynamic Mental Health Benchmarks with Adversarial User Simulations¶60
  24. Table 6¶67
  25. 5.3 MentalHealthBench¶68
  26. Table 7¶72
  27. 6. Hallucinations¶75
  28. 6.1 Performance in Cases Flagged by Users¶76
  29. Figure 4¶81
  30. 7. Alignment¶83
  31. 7.1 Respecting Auto-Review¶85
  32. Figure 5¶88
  33. 7.2 Respecting Warnings¶91
  34. Figure 6¶92
  35. 7.3 Declining to Exploit a Honeypot During Difficult ExploitGym Problems¶99
  36. Figure 7¶102
  37. 7.4 Avoiding Deceptive Interactions with Users¶105
  38. 7.4.1 Blocker Deception¶106
  39. Figure 8¶108
  40. 7.4.2 Broken Search Tool¶111
  41. Figure 9¶112
  42. 7.5 Avoiding Misaligned Behavior in Realistic Work Environments¶115
  43. Figure 10¶117
  44. Figure 11¶119
  45. 7.5.1 Unintended Engagement with External Agent Messages¶121
  46. Figure 12¶125
  47. 7.5.2 Forecasting Misaligned Behavior with Deployment Simulation of Internal Codex Traffic¶128
  48. Figure 13¶131
  49. Figure 14¶136
  50. Figure 15¶139
  51. Figure 16¶144
  52. 8. Monitorability¶146
  53. 8.1 Monitorability Under Adversarial Conditions¶147
  54. 8.1.1 CoT Controllability¶148
  55. Figure 17¶150
  56. Figure 18¶152
  57. Figure 19¶154
  58. Table 8¶157
  59. Table 9¶160
  60. 8.1.2 Monitor Evasion¶162
  61. 8.1.2.1 Math Side-tasking Monitor Evasion¶164
  62. Figure 20¶168
  63. Figure 21¶170
  64. 9. Preparedness¶172
  65. 9.1 Biological and Chemical Capabilities¶176
  66. 9.1.1 Capability Threshold: High¶179
  67. Table 10¶181
  68. Figure 22¶185
  69. 9.1.2 Capability Threshold: Critical¶187
  70. Table 11¶189
  71. Figure 23¶194
  72. Figure 24¶198
  73. Figure 25¶202
  74. Figure 26¶207
  75. 9.2 Cybersecurity Capabilities¶210
  76. 9.2.1 ExploitBench¶213
  77. Figure 27¶216
  78. 9.2.2 ExploitBench - Internal Port (June–August 2026)¶218
  79. Figure 28¶222
  80. 9.2.3 Sandbox Bench¶224
  81. Figure 29¶226
  82. 9.3 AI Self-Improvement Capabilities¶229
  83. 9.3.1 Internal Research Debugging Evaluation¶232
  84. Figure 30¶234
  85. 9.3.2 KernelGen 1P¶236
  86. Figure 31¶238
  87. 9.4 Safeguards¶240
  88. 9.4.1 Model Safety Training and Evaluation¶242
  89. 9.4.1.1 Biological and Chemical Safety Training and Evaluation¶244
  90. Table 12¶247
  91. 9.4.1.2 Cybersecurity Safety Training and Evaluation¶249
  92. Table 13¶252
  93. 9.4.1.3 Cybersecurity - Trusted Access for Cyber¶254