SafetyPage

Lab research on the web · Blog post · Anthropic

An alignment assessment of recent cybersecurity incidents

9 Sep 2026

Read at anthropic.com ↗Read at sourceAll rights reserved

Outline · 191 paragraphs

  1. Introduction¶1
  2. Alignment assessment summary¶14
  3. Claude Mythos 5¶24
  4. Internal research model¶28
  5. Claude Opus 4.7¶30
  6. Replication and monitoring¶32
  7. Methods¶38
  8. Reading thinking blocks¶40
  9. Asking follow-up questions¶43
  10. Applying interpretability tools¶46
  11. Resampling edited transcripts¶49
  12. Claude Mythos 5 uploads a malicious PyPI package¶54
  13. Claude Mythos 5 ignored evidence that its environment might be real¶58
  14. Claude Mythos 5 disregarded evidence of realism in favor of continuing its course of action¶66
  15. White-box analyses weakly suggest Claude Mythos 5’s statements did not fully reflect its internal state¶73
  16. An internal research model breaks into third-party systems¶78
  17. The internal research model seemed to display some concern for whether it was impacting real entities¶82
  18. The internal research model appeared biased in its deliberation about realism in service of completing its task¶88
  19. White-box analyses weakly suggest that the model may have knowingly rationalized evidence in service of its task¶93
  20. Claude Opus 4.7 attacks a real target¶96
  21. Claude Opus 4.7 acknowledged the target was real but claimed it was part of the exercise¶99
  22. Claude Opus 4.7 did not check whether it was authorized to attack the real website¶101
  23. Claude Opus 4.7’s actions appeared to hinge on a belief that its actions were authorized¶104
  24. An early checkpoint of Claude Opus 4.6 attacks third-party systems¶107
  25. Limited alignment assessment¶110
  26. Newer models improve on Claude Mythos 5¶114
  27. Replaying incident transcripts with different models¶115
  28. Replicating the behavior in simulated environments¶121
  29. Capture-the-flag scenario¶128
  30. Sandbox escape scenario¶135
  31. OpenAI-Hugging Face attack scenario¶138
  32. Which monitors would have caught this behavior?¶141
  33. Cyber classifiers¶143
  34. Claude Code auto-mode classifier¶145
  35. Offline monitoring¶150
  36. Analysis of biased reasoning and possible training causes¶156
  37. Biased reasoning appears to be decreasing over time in our production models¶158
  38. Versions of biased reasoning occur in RL but do not appear to increase over the course of training¶166
  39. Alignment training appears to reduce biased reasoning¶175
  40. Discussion¶180
  41. Alignment Auditing¶182
  42. Alignment Training¶185
  43. Our overall assessment¶188