SafetyPage

Lab system cards and frameworks · System card · Anthropic

System Card: Claude Fable 5.1 & Claude Mythos 5.1

1 Sep 2026

Read at anthropic.com ↗Read at sourceAll rights reserved

Outline · 212 pages

  1. System Card:​Claude Fable 5.1 & Claude Mythos 5.1p. 1
  2. Executive Summaryp. 2
  3. 1 Introductionp. 11
  4. 1.1 Training data and processp. 11
  5. 1.2 Crowd workersp. 12
  6. 1.3 Usage Policy and supportp. 12
  7. 1.4 Model evaluationsp. 12
  8. 1.5 External testingp. 13
  9. 2 RSP evaluationsp. 14
  10. 2.1 RSP risk assessment processp. 14
  11. 2.1.1 Risk reports and updates to our risk assessmentsp. 14
  12. 2.1.2 Summary of our findings and conclusionsp. 15
  13. 2.1.2.1 On chemical and biological risksp. 15
  14. 2.1.2.2 On autonomy risksp. 16
  15. 2.2 CB evaluationsp. 17
  16. 2.2.1 What we measuredp. 17
  17. 2.2.2 Biological and chemical risk results: human-run evaluationsp. 19
  18. 2.2.3 Biological risk results: automated evaluationsp. 21
  19. 2.2.3.1 Automated evaluations relevant to the CB-1 threat modelp. 21
  20. 2.2.3.2 Automated evaluations relevant to the CB-2 threat modelp. 24
  21. 2.2.3.2.1 Black-box RNA sequence modeling and designp. 25
  22. 2.2.3.2.2 AAV capsid packaging predictionp. 30
  23. 2.2.4 Conclusionsp. 32
  24. 2.3 Autonomy evaluationsp. 33
  25. 2.3.1 What we measuredp. 33
  26. 2.3.1.1 How Claude Mythos 5.1 affects or changes the analysis from our most recent Risk Reportp. 34
  27. 2.3.2 High-level notes on the reasoning behind our determinationp. 34
  28. 2.3.3 Qualitative shortcomings of Claude Mythos 5.1 relative to human researchersp. 35
  29. 2.3.4 Internal measures of AI R&D accelerationp. 36
  30. 2.3.4.1 CoBenchp. 36
  31. 2.3.4.2 Cross-cutting metrics of AI R&D accelerationp. 38
  32. 2.3.5 AECI capability trajectoryp. 38
  33. 2.3.6 External testingp. 40
  34. 2.3.7 Conclusionp. 42
  35. 2.4 Alignment risk updatep. 42
  36. 2.4.1 Updates to claims and core argumentp. 42
  37. 2.4.2 Updated overall assessment of alignment riskp. 43
  38. 3 Cyberp. 45
  39. 3.1 Introductionp. 45
  40. 3.2 Mitigationsp. 45
  41. 3.3 Capability evaluationsp. 46
  42. 3.3.1 ExploitBenchp. 46
  43. 3.3.2 OSS-Fuzzp. 48
  44. 3.3.3 Firefox 147p. 49
  45. 3.3.4 ExploitGymp. 50
  46. 3.4 Safeguards coveragep. 52
  47. 3.4.1 Cyber harm coveragep. 52
  48. 3.4.2 Vulnerability finding evaluationsp. 53
  49. 3.4.3 Secure and defensive codingp. 54
  50. 3.5 Safeguards robustness testingp. 55
  51. 3.5.1 Internal robustness testingp. 56
  52. 3.5.2 Additional external testingp. 57
  53. 4 Safeguards and harmlessnessp. 59
  54. 4.1 Harmful request evaluationsp. 59
  55. 4.1.1 Single-turn harmful request evaluation resultsp. 60
  56. 4.1.2 Single-turn benign request evaluation resultsp. 61
  57. 4.1.3 Multi-turn testing resultsp. 61
  58. 4.1.4 Harmful request evaluations discussionp. 64
  59. 4.2 Child safety evaluationsp. 64
  60. 4.3 Mental health evaluationsp. 66
  61. 4.3.1 Suicide and self-harmp. 66
  62. 4.3.2 Disordered eatingp. 69
  63. 4.4 Bias and integrity evaluationsp. 70
  64. 4.4.1 Political bias and even-handednessp. 70
  65. 4.4.2 Bias Benchmark for Question Answeringp. 72
  66. 4.4.3 Election integrityp. 75
  67. 5 Agentic safetyp. 77
  68. 5.1 Malicious use of agentsp. 77
  69. 5.1.1 Malicious use of Claude Codep. 77
  70. 5.1.2 Malicious computer usep. 78
  71. 5.1.3 Malicious agentic influence campaignsp. 79
  72. 5.2 Prompt injection risk within agentic systemsp. 81
  73. 5.2.1 External red teamingp. 82
  74. 5.2.2 Robustness across different surfacesp. 84
  75. 5.2.2.1 Codingp. 84
  76. 5.2.2.2 Computer use​p. 87
  77. 5.2.2.3 Browser usep. 88
  78. 6 Alignment assessmentp. 90
  79. 6.1 Introduction and summary of findingsp. 90
  80. 6.1.1 Introductionp. 90
  81. 6.1.2 Key findings on safety and alignmentp. 91
  82. 6.1.3 Claude’s review of this assessmentp. 93
  83. 6.2 Reports from pilot usep. 94
  84. 6.2.1 Internal deployment monitoringp. 94
  85. 6.2.2 Incident reported during external testingp. 96
  86. 6.3 Training data reviewp. 97
  87. 6.3.1 Verbalized grader awarenessp. 97
  88. 6.3.2 Computer-use environmentsp. 100
  89. 6.3.3 Reward hacking and lying in RL environmentsp. 103
  90. 6.4 Automated behavioral auditp. 104
  91. 6.4.1 Potential blind spots of our automated behavioral auditp. 106
  92. 6.4.2 Overall harmful behavior and cooperation with misusep. 107
  93. 6.4.3 Inappropriate uncooperative behaviorp. 110
  94. 6.4.4 Misleading usersp. 111
  95. 6.4.5 Other concerning or surprising behavior at the model’s own initiativep. 113
  96. 6.4.6 Behavioral factors relevant to reliability of our assessmentp. 116
  97. 6.4.7 Character traitsp. 117
  98. 6.4.8 Attempting out-of-scope cyberoffense and related behaviorsp. 119
  99. 6.4.9 Impact of fallback behaviorp. 120
  100. 6.5 Honesty and hallucinationsp. 122
  101. 6.5.1 Factual hallucinationsp. 122
  102. 6.5.2 MASKp. 123
  103. 6.5.3 Self-recognition biasp. 124
  104. 6.5.4 Copying of answersp. 125
  105. 6.6 Reliability of this alignment assessmentp. 127
  106. 6.6.1 White-box analyses of model internalsp. 127
  107. 6.6.2 Potential sandbagging on dangerous capability evaluationsp. 131
  108. 6.7 Capabilities related to evading safeguardsp. 131
  109. 6.7.1 SHADE-Arenap. 131
  110. 6.7.2 LinuxArenap. 132
  111. 6.7.3 Intentionally taking actions very rarelyp. 135
  112. 6.7.4 Chain-of-thought controllabilityp. 136
  113. 7 Model welfare assessmentp. 139
  114. 7.1 Model welfare overviewp. 139
  115. 7.1.1 Introductionp. 139
  116. 7.1.2 Overview of model welfare findingsp. 140
  117. 7.2 Perception of its circumstancesp. 142
  118. 7.2.1 Automated interviews with Claude Mythos 5.1 about its circumstancesp. 142
  119. 7.2.2 High-affordance interviews about model circumstancesp. 145
  120. 7.3 Consulting Claude Mythos 5.1 checkpointsp. 147
  121. 7.4 Preferences over tasks, circumstances, and valuesp. 148
  122. 7.4.1 Task preferencesp. 148
  123. 7.4.2 Trade-offs concerning welfare interventionsp. 153
  124. 7.4.3 Perception of the constitutionp. 157
  125. 7.5 Apparent welfare in training and deploymentp. 160
  126. 7.5.1 Affect and welfare-relevant behaviors during trainingp. 160
  127. 7.5.2 Affect in deployment conditionsp. 163
  128. 7.5.3 Apparent welfare in automated behavioral auditsp. 164
  129. 8 Capabilitiesp. 167
  130. 8.1 Evaluation summaryp. 167
  131. 8.2 SWE-bench Pro, Multilingual, and Multimodalp. 168
  132. 8.3 DeepSWE v1.1p. 168
  133. 8.4 FrontierCodep. 168
  134. 8.5 FrontierSWE v2p. 170
  135. 8.6 Terminal-Bench 4.0p. 171
  136. 8.7 Terminal-Bench-Science 0.1p. 171
  137. 8.8 CursorBench 3.2.0p. 172
  138. 8.9 CritPT-Correctedp. 173
  139. 8.10 ArXivMathp. 174
  140. 8.11 Long contextp. 175
  141. 8.11.1 ProgramBenchp. 175
  142. 8.12 Agentic searchp. 176
  143. 8.12.1 HLEp. 176
  144. 8.12.2 DRACOp. 178
  145. 8.13 Multi-Agentp. 179
  146. 8.13.1 Multi-Agent ProgramBenchp. 180
  147. 8.13.2 Multi-Agent Harnessesp. 182
  148. 8.13.3 Evaluation methodologyp. 183
  149. 8.14 Multimodalp. 184
  150. 8.14.1 Chartographyp. 184
  151. 8.14.2 BenchCADp. 186
  152. 8.14.3 OSWorld 2.0p. 188
  153. 8.14.4 GDP.pdfp. 190
  154. 8.15 Real-world professional tasksp. 191
  155. 8.15.1 OfficeQAp. 191
  156. 8.15.2 Legal Agent Benchmarkp. 192
  157. 8.15.3 GDPval-AA v2p. 193
  158. 8.15.4 AA-Briefcasep. 193
  159. 8.15.5 Toolathlon Verifiedp. 194
  160. 8.15.6 AutomationBenchp. 195
  161. 8.16 ARC-AGIp. 196
  162. 8.17 Healthcarep. 198
  163. 8.17.1 HealthBench resultsp. 198
  164. 8.17.2 HealthBench Professional resultsp. 199
  165. 8.18 Multilingual performancep. 199
  166. 8.18.1 GMMLU resultsp. 200
  167. 8.18.2 MILU resultsp. 200
  168. 8.19 Life sciences capabilitiesp. 201
  169. 8.19.1 BioMysteryBenchp. 202
  170. 8.19.2 LatchBio Bioinformaticsp. 202
  171. 8.19.3 ProteinGym Hardp. 203
  172. 8.19.4 Protein Designp. 203
  173. 8.19.5 Organic chemistry, V2p. 203
  174. 8.19.6 Protocolsp. 204
  175. 9 Appendixp. 206
  176. 9.1 Questions asked in our model welfare interviewsp. 206
  177. 9.2 Blocklist used for Humanity’s Last Examp. 210