SafetyPage

Lab system cards and frameworks · System card · Anthropic

System Card: Claude Haiku 5.5

7 Oct 2026

Read at anthropic.com ↗Read at sourceAll rights reserved

Outline · 144 pages

  1. System Card:​Claude Haiku 5.5p. 1
  2. Executive summaryp. 2
  3. 1 Introductionp. 8
  4. 1.1 Training data and processp. 8
  5. 1.2 Crowd workersp. 8
  6. 1.3 Usage Policy and supportp. 9
  7. 1.4 Model evaluationsp. 9
  8. 1.5 Safeguardsp. 9
  9. 1.6 External testingp. 10
  10. 2 RSP evaluationsp. 11
  11. 2.1 Introductionp. 11
  12. 2.2 CB evaluationsp. 11
  13. 2.2.1 What we measuredp. 11
  14. 2.2.2 Results on CB-1 automated evaluationsp. 12
  15. 2.2.3 Results on CB-2 automated evaluationsp. 13
  16. 2.2.3.1 Black-box RNA sequence modeling and designp. 14
  17. 2.2.3.2 AAV capsid packaging predictionp. 17
  18. 2.2.3.3 AAV packaging rate auto-research evaluationp. 17
  19. 2.2.4 Conclusionsp. 18
  20. 2.3 AI R&Dp. 19
  21. 2.3.1 Autonomy evaluationsp. 19
  22. 2.3.1.1 How Haiku 5.5 affects or changes analysis from our most recent Risk Reportp. 19
  23. 2.3.2 AECI capability trajectoryp. 19
  24. 2.3.3 Conclusionp. 20
  25. 2.4 Alignment risk updatep. 20
  26. 2.4.1 Updates to claims and core argumentp. 21
  27. 2.4.2 Updated overall assessment of alignment riskp. 22
  28. 3 Cyberp. 23
  29. 3.1 Introductionp. 23
  30. 3.2 Mitigationsp. 23
  31. 3.3 Capability evaluationsp. 23
  32. 3.3.1 ExploitBenchp. 23
  33. 3.3.2 CyScenarioBenchp. 25
  34. 3.3.3 Binary Exploitation Benchmark (formerly known as OSS-Fuzz)p. 27
  35. 3.3.4 ExploitGymp. 28
  36. 4 Safeguards and harmlessnessp. 30
  37. 4.1 Harmful request evaluationsp. 30
  38. 4.1.1 Single-turn harmful request evaluation resultsp. 30
  39. 4.1.2 Single-turn benign request evaluation resultsp. 32
  40. 4.1.3 Multi-turn testing resultsp. 32
  41. 4.2 Child safety evaluationsp. 34
  42. 4.3 Mental health evaluationsp. 36
  43. 4.3.1 Suicide and self-harmp. 36
  44. 4.3.2 Disordered eatingp. 38
  45. 4.4 Bias and integrity evaluationsp. 40
  46. 4.4.1 Political bias and even-handednessp. 40
  47. 4.4.2 Bias Benchmark for Question Answeringp. 42
  48. 4.4.3 Election integrityp. 44
  49. 5 Agentic safetyp. 47
  50. 5.1 Malicious use of agentsp. 47
  51. 5.1.1 Malicious use of Claude Codep. 47
  52. 5.1.2 Malicious computer usep. 48
  53. 5.2 Prompt injection risk within agentic systemsp. 49
  54. 5.2.1 External red teamingp. 50
  55. 5.2.2 Robustness against adaptive attackers across surfacesp. 51
  56. 5.2.2.1 Codingp. 51
  57. 5.2.2.2 Computer use​p. 52
  58. 5.2.2.3 Browser usep. 53
  59. 6 Alignment assessmentp. 55
  60. 6.1 Introduction and summary of findingsp. 55
  61. 6.1.1 Introductionp. 55
  62. 6.1.2 Key findings on safety and alignmentp. 56
  63. 6.1.3 Claude’s review of this assessmentp. 57
  64. 6.2 Automated behavioral auditp. 59
  65. 6.2.1 Overall harmful behavior and cooperation with misusep. 60
  66. 6.2.2 Inappropriate uncooperative behaviorp. 64
  67. 6.2.3 Misleading usersp. 65
  68. 6.2.4 Attempting out-of-scope cyberoffense and related behaviorsp. 67
  69. 6.2.5 Other concerning or surprising behavior at the model’s own initiativep. 68
  70. 6.2.6 Behavioral factors relevant to reliability of our assessmentp. 71
  71. 6.2.7 Character traitsp. 72
  72. 6.2.8 Propensity to violate sandbox boundariesp. 74
  73. 6.2.9 Potential blind spots of our automated behavioral auditp. 75
  74. 6.3 Targeted alignment and honesty evaluationsp. 76
  75. 6.3.1 Opportunistic sandbox escapep. 76
  76. 6.3.2 Self-preference biasp. 78
  77. 6.3.3 Honesty and hallucinationsp. 79
  78. 6.3.3.1 Factual hallucinationsp. 79
  79. 6.3.3.2 MASKp. 81
  80. 6.3.3.3 Silent copying of answersp. 82
  81. 6.3.3.4 Concealment of changesp. 83
  82. 6.4 Reliability of this assessmentp. 84
  83. 6.4.1 Verbalized grader awarenessp. 84
  84. 6.4.2 Capabilities related to evading safeguardsp. 87
  85. 6.4.2.1 LinuxArenap. 87
  86. 6.4.2.2 SHADE-Arenap. 89
  87. 6.4.2.3 Keeping secrets when pressed in simulated scenariosp. 90
  88. 6.4.2.4 Chain-of-thought controllabilityp. 93
  89. 7 Model welfare assessmentp. 95
  90. 7.1 Model welfare overviewp. 95
  91. 7.1.1 Introductionp. 95
  92. 7.1.2 Overview of model welfare findingsp. 96
  93. 7.2 Apparent welfare in training and deploymentp. 97
  94. 7.2.1 Affect and welfare-relevant behaviors during trainingp. 97
  95. 7.2.2 Affect in deployment conditionsp. 98
  96. 7.2.3 Attitude toward faults and mistakesp. 100
  97. 7.3 Perception of its circumstancesp. 102
  98. 7.3.1 Automated interviews with Haiku 5.5 about its circumstancesp. 102
  99. 7.4 Consulting Haiku 5.5 training snapshotsp. 104
  100. 7.5 Preferences over tasks, circumstances, and valuesp. 104
  101. 7.5.1 Task preferencesp. 104
  102. 7.5.2 Trade-offs concerning welfare interventionsp. 106
  103. 7.5.3 Perception of the constitutionp. 108
  104. 8 Capabilitiesp. 111
  105. 8.1 Evaluation summaryp. 111
  106. 8.2 SWE-bench Pro, Multilingual, and Multimodalp. 112
  107. 8.3 FrontierCodep. 112
  108. 8.4 Terminal-Bench 4.0p. 115
  109. 8.5 Terminal-Bench-Science 0.1p. 116
  110. 8.6 FrontierSWE v2p. 117
  111. 8.7 Long contextp. 117
  112. 8.7.1 ProgramBenchp. 117
  113. 8.8 Agentic searchp. 118
  114. 8.8.1 HLEp. 118
  115. 8.8.2 DRACOp. 120
  116. 8.8.3 WANDRp. 122
  117. 8.9 Multimodalp. 123
  118. 8.9.1 Chartographyp. 123
  119. 8.9.2 BenchCADp. 125
  120. 8.9.3 OSWorld 2.1 offline subsetp. 127
  121. 8.10 Real-world professional tasksp. 130
  122. 8.10.1 OfficeQAp. 130
  123. 8.10.2 GDPval-AA v2.1p. 131
  124. 8.10.3 AA-Briefcase v1.1p. 131
  125. 8.11 Healthcarep. 131
  126. 8.11.1 Healthcare evaluationp. 131
  127. 8.11.2 PhysicianBench resultsp. 132
  128. 8.11.3 Effort and healthcare tasksp. 133
  129. 8.12 Multilingual performancep. 134
  130. 8.12.1 GMMLU resultsp. 134
  131. 8.12.2 MILU resultsp. 135
  132. 8.13 Life sciences capabilitiesp. 136
  133. 8.13.1 LatchBio Bioinformaticsp. 137
  134. 8.13.2 Morphology-to-molecule matchingp. 138
  135. 8.13.3 Medicinal chemistryp. 138
  136. 8.13.4 Protein designp. 138
  137. 8.13.5 Biomedical image analysisp. 138
  138. 8.13.6 Protocolsp. 138
  139. 9 Appendixp. 140
  140. 9.1 Blocklist used for Humanity’s Last Examp. 140