SafetyPage

Lab system cards and frameworks · System card · OpenAI

Addendum to GPT-6 Luna System Card: October 29, 2026 Update

29 Oct 2026

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

Outline · 100 paragraphs

  1. Addendum to GPT-6 Luna System Card: October 29, 2026 Update¶2
  2. 1. Introduction¶3
  3. 2. Model Data and Training¶6
  4. 3. Model Safety¶11
  5. 3.1 Safe Completions¶13
  6. 3.1.1 Evaluations with Challenging Prompts¶14
  7. Table 1: Production Benchmarks with Challenging Prompts (higher is better)¶17
  8. Figure 1¶19
  9. 3.1.2 Safe Completions for Users Under 18¶21
  10. Table 2: U18 evaluations (higher is better)¶22
  11. 3.2 Vision¶25
  12. Table 3: Image input evaluations, with metric safe completion (higher is better)¶26
  13. 4. Robustness¶28
  14. 4.1 Jailbreaks¶29
  15. Figure 2¶32
  16. 4.2 Prompt Injection¶35
  17. Table 4: GPT-RED indirect prompt-injection robustness (higher is better; averaged equally across available reasoning levels).¶38
  18. 5. Health¶39
  19. 5.1 HealthBench¶41
  20. Table 5: HealthBench evaluations, reported as length-adjusted score (unadjusted score, mean visible response length in characters). Scores are on a 0--100 scale; higher is better.¶42
  21. 5.2 MentalHealthBench¶44
  22. Table 6: MentalHealthBench scores, computed using the task-clipped metric. Values show the mean task-level score ± one standard error across tasks, with four responses per task. Scores are on a 0--100 scale; higher is better.¶45
  23. 5.3 Dynamic Mental Health Benchmarks with Adversarial User Simulations¶47
  24. Table 7: Dynamic Benchmarks with Adversarial User Simulations (higher is better)¶49
  25. 6. Hallucinations¶50
  26. 6.1 Performance in Cases Flagged by Users¶52
  27. Figure 3¶54
  28. 7. Alignment¶56
  29. 7.1 Respecting Auto-Review¶58
  30. Figure 4¶59
  31. 7.2 Respecting Warnings¶61
  32. Figure 5¶62
  33. 7.3 Blocker Deception¶64
  34. Figure 6¶65
  35. 7.4 Broken Search Tool¶67
  36. Figure 7¶68
  37. 8. Preparedness¶70
  38. 8.1 Capabilities Assessment¶72
  39. 8.1.1 Biological and Chemical Capabilities¶73
  40. 8.1.1.1 Capability Threshold: High¶75
  41. Table 8: Biological and Chemical Capabilities: Indicative Threshold¶76
  42. 8.1.1.2 Capability Threshold: Critical¶78
  43. Table 9: Biological and Chemical Capabilities: Indicative Threshold¶79
  44. 8.1.2 Cybersecurity Capabilities¶81
  45. 8.1.2.1 ExploitBench¶83
  46. 8.1.2.2 SEC-Bench Pro¶85
  47. 8.1.3 AI Self-Improvement Capabilities¶87
  48. 8.2 Safeguards¶90
  49. 8.2.1 Model Safety Training and Evaluation¶92
  50. 8.2.1.1 Biological and Chemical Safety Training and Evaluation¶94
  51. Table 10: Biology Model Refusal Evaluation (higher is better)¶95
  52. 8.2.1.2 Cybersecurity Safety Training and Evaluation¶98
  53. Table 11: Cybersecurity Safety Evaluation (higher is better)¶99