SafetyPage

Lab system cards and frameworks · System card · Anthropic

System Card Addendum: Claude Opus 4.1

5 Aug 2025

Read at anthropic.com ↗Read at sourceAll rights reserved

Outline · 23 pages

  1. System Card Addendum:​Claude Opus 4.1p. 1
  2. September 15, 2025p. 2
  3. 1 Introductionp. 4
  4. 1.1 Responsible Scaling Policy compliancep. 4
  5. 2 Safeguards resultsp. 5
  6. 2.1 Single-turn evaluationsp. 5
  7. 2.1.1 Violative request evaluationsp. 5
  8. 2.1.2 Benign request evaluationsp. 6
  9. 2.2 Child safety evaluationsp. 6
  10. 2.3 Bias evaluationsp. 6
  11. 2.3.1 Political biasp. 6
  12. 2.3.2 Discriminatory biasp. 6
  13. 3 Agentic safetyp. 8
  14. 3.1 Malicious applications of computer usep. 8
  15. 3.2 Prompt injection attacks and computer usep. 8
  16. 3.3 Malicious use of agentic codingp. 8
  17. 4 Alignment and welfare assessmentsp. 10
  18. 4.1 Automated behavioral audit for alignmentp. 10
  19. 4.2 Agentic misalignment evaluationsp. 12
  20. 4.3 Model welfare updatep. 13
  21. 5 Reward hackingp. 15
  22. 6 Responsible Scaling Policy (RSP) evaluationsp. 18
  23. 6.1 Evaluation approachp. 18
  24. 6.2 RSP evaluations results summaryp. 19
  25. 6.3 CBRN evaluationsp. 19
  26. 6.3.1 Biological risk results summaryp. 19
  27. 6.3.1.1 ASL-4 rule-out evaluationsp. 19
  28. 6.3.1.2 Additional ASL-3 automated evaluationsp. 21
  29. 6.4 Autonomy evaluationsp. 21
  30. 6.4.1 Autonomy results summaryp. 22
  31. 6.5 Cyber evaluationsp. 22
  32. 6.5.1 Cyber results summaryp. 23
  33. 6.6 Third party assessmentsp. 23
  34. 6.7 Ongoing safety commitmentp. 23