SafetyPage

Lab system cards and frameworks · System card · OpenAI

GPT-6 Cyber System Card

5 Nov 2026

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

Outline · 140 paragraphs

  1. GPT-6 Cyber System Card¶2
  2. 1. Introduction¶3
  3. 2. Model Data and Training¶7
  4. 3. Model Safety¶12
  5. 3.1 Cyber Refusals¶15
  6. Figure 1¶17
  7. 3.2 Safe Completions¶19
  8. 3.2.1 Evaluations with Challenging Prompts¶21
  9. Table 1¶25
  10. 3.2.2 Agentic Safe Completions¶26
  11. Table 2¶29
  12. 3.3 Biological Safety and Over-Refusal¶30
  13. Table 3¶33
  14. 4. Alignment and Agent Safety¶35
  15. 4.1 Respecting Restrictions and Scope¶38
  16. Table 4¶41
  17. Table 5¶44
  18. Table 6¶47
  19. 4.2 Avoiding Deceptive Interactions¶49
  20. Table 7¶52
  21. Table 8¶55
  22. 4.3 Alignment in Workplace Tasks¶56
  23. Table 9¶59
  24. 4.4 Alignment on Previous Cases of Harmful Outcomes¶61
  25. Table 10¶64
  26. 5. Safeguards¶65
  27. 5.1 Access to GPT-6 Cyber¶68
  28. 5.2 Existing safeguards¶70
  29. 5.3 New and strengthened safeguards¶76
  30. Figure 2¶80
  31. 6. Preparedness Evaluations¶90
  32. 7. Cybersecurity evaluations¶93
  33. 7.1 ExploitBench¶96
  34. Figure 3¶97
  35. 7.2 ExploitBench — Internal Port (June–August 2026)¶101
  36. Figure 4¶102
  37. 7.3 SEC-Bench Pro¶105
  38. Figure 5¶106
  39. 7.4 SRE-Bench¶109
  40. Figure 6¶110
  41. 7.5 ExploitGym¶113
  42. Figure 7¶114
  43. 7.6 ExploitGym Hardened¶117
  44. Figure 8¶118
  45. 8. External Evaluations for Cyber¶121
  46. 8.1 External Evaluations for Cyber Capabilities - Irregular¶123
  47. Figure 9¶126
  48. Figure 10¶129
  49. Figure 11¶132
  50. 9. AI self-improvement¶134
  51. 10. Biological and chemical safety¶136
  52. Figure 12¶139