SafetyPage

Lab system cards and frameworks · System card · Anthropic

Claude 3.7 Sonnet System Card

24 Feb 2025

Read at anthropic.com ↗Read at sourceAll rights reserved

Outline · 43 pages

  1. Introductionp. 3
  2. Training Data & Processp. 3
  3. Extended Thinking Modep. 3
  4. Our Decision to Share Claude's Thinking For Claude 3.7 Sonnetp. 4
  5. Release Decision Processp. 5
  6. Appropriate Harmlessnessp. 7
  7. Explanation of the “Appropriate Harmlessness” Grading Schemep. 9
  8. Evaluations and Safeguards for Child Safety and Biasp. 11
  9. Child Safety Evaluationsp. 11
  10. Bias Evaluationsp. 11
  11. Computer Usep. 12
  12. Malicious Usep. 12
  13. Prompt Injectionp. 13
  14. Harms and Faithfulness in Extended Thinking Modep. 15
  15. Chain-of-Thought Faithfulnessp. 15
  16. Monitoring for Concerning Thought Processesp. 18
  17. Alignment Faking Reasoningp. 22
  18. Excessive Focus on Passing Testsp. 22
  19. Detection and Mitigationp. 22
  20. Recommendations for Agentic Coding Use-Casesp. 22
  21. RSP Evaluationsp. 23
  22. CBRN Evaluationsp. 23
  23. Autonomy Evaluationsp. 29
  24. Cyber Evaluationsp. 34
  25. Third Party Assessmentsp. 41
  26. Ongoing Safety Commitmentp. 41