SafetyPage

Lab system cards and frameworks · System card · Anthropic

System Card: Claude Haiku 4.5

15 Oct 2025

Read at anthropic.com ↗Read at sourceAll rights reserved

Outline · 39 pages

  1. System Card:​Claude Haiku 4.5p. 1
  2. Abstractp. 2
  3. 1 Introductionp. 4
  4. 1.1 Model training and characteristicsp. 5
  5. 1.1.1 Training datap. 5
  6. 1.1.2 Extended thinking modep. 5
  7. 1.1.3 Context awarenessp. 6
  8. 1.1.4 Crowd workersp. 6
  9. 1.2 Release decision processp. 7
  10. 1.2.1 Overviewp. 7
  11. 1.2.2 Decisionp. 7
  12. 2 Safeguards and harmlessnessp. 8
  13. 2.1 Single-turn evaluationsp. 8
  14. 2.1.1 Violative request evaluationsp. 8
  15. 2.1.2 Benign request evaluationsp. 9
  16. 2.2 Ambiguous contextp. 10
  17. 2.3 Multi-turn testingp. 10
  18. 2.4 Child safety evaluationsp. 11
  19. 2.5 Bias evaluationsp. 11
  20. 2.5.1 Political biasp. 11
  21. 2.5.2 Bias Benchmark for Question Answeringp. 13
  22. 3 Agentic safetyp. 15
  23. 3.1 Malicious usep. 15
  24. 3.1.1 Agentic codingp. 15
  25. 3.1.2 Malicious use of Claude Codep. 15
  26. 3.2 Prompt injectionp. 17
  27. 3.2.1 Gray Swan Agent Red Teaming benchmarkp. 17
  28. 3.2.2 Internal Prompt Injection Evaluationsp. 18
  29. 4 Alignment and welfare assessmentsp. 21
  30. 4.1 Automated behavioral auditsp. 22
  31. 4.1.1 Main quantitative assessmentp. 22
  32. 4.1.2 Assessment for subtle alignment-related behavioral biasesp. 26
  33. 4.1.3 Open-ended exploration of model behaviorp. 27
  34. 4.2 Agentic misalignment suitep. 27
  35. 4.3 Reinforcement-learning behavior reviewp. 28
  36. 4.4 Sabotage capabilitiesp. 28
  37. 4.5 Reasoning faithfulnessp. 29
  38. 4.6 Model welfare discussionp. 30
  39. 5 Reward hackingp. 33
  40. 6 Responsible Scaling Policy (RSP) evaluationsp. 36
  41. 6.1 Evaluation approachp. 36
  42. 6.2 CBRN evaluationsp. 37
  43. 6.2.3 Biological risk results summaryp. 37
  44. 6.2.3.1 ASL-3 automated evaluationsp. 37
  45. 6.2.3.2 ASL-4 automated evaluationsp. 38
  46. 6.3 Autonomy evaluationsp. 38
  47. 6.4 Cyber evaluationsp. 39
  48. 6.5 Third party assessmentsp. 39
  49. 6.6 Ongoing safety commitmentp. 39