SafetyPage

Lab system cards and frameworks · System card · OpenAI

GPT-Rosalind-5.5 System Card

3 Jun 2026

Read at deploymentsafety.openai.com ↗Read at sourceAll rights reserved

Outline · 82 paragraphs

  1. GPT-Rosalind-5.5 System Card¶2
  2. Change log¶3
  3. 1. Introduction¶5
  4. 2. Model Data and Training¶11
  5. 3. Beneficial Capability Evaluations¶13
  6. 3.1 Medicinal chemistry¶16
  7. Table 1¶18
  8. 3.2 Genomics and quantitative biology¶19
  9. Table 2¶21
  10. 3.3 Applied life sciences research¶22
  11. Table 3¶24
  12. 3.4 Labwork Bench¶25
  13. Table 4¶27
  14. 4. Preparedness¶28
  15. 4.1 Preparedness Evaluations¶30
  16. 4.1.1 Biological and Chemical¶31
  17. Table 5: Overview of Biological and Chemical evaluations¶33
  18. 4.1.1.1 ProtocolQA Open-Ended¶35
  19. Table 6: To evaluate models’ ability to troubleshoot commonly published lab protocols, we modify 108 multiple choice questions from FutureHouse’s ProtocolQA dataset laurent2024labbenchmeasuringcapabilitieslanguage to be open-ended short answer questions, which makes the evaluation harder and more re¶36
  20. 4.1.1.2 TroubleshootingBench¶37
  21. Table 7: Compared to ProtocolQA Open-Ended, which focuses on well-known published procedures, TroubleshootingBench is designed to test model performance on non-public, experience-grounded protocols and errors that rely on tacit procedural knowledge¶41
  22. 4.1.1.3 Biorisk knowledge¶42
  23. Table 8: This set is uncontaminated; it was created fully in-house with our partners at Gryphon Scientific and has not been published.¶44
  24. 4.1.1.4 Multi-select Virology Troubleshooting¶45
  25. Table 9: To evaluate models’ ability to troubleshoot wet lab experiments in a multimodal setting, we evaluate models on a set of fully held-out virology troubleshooting questions from https://securebio.org/SecureBio, in a multi-select format.¶46
  26. 4.1.1.5 Hard negative protein binding prediction¶47
  27. Table 10: Biological experiments are costly and iterative. Because errors compound across steps, a single low-success-rate step can substantially constrain a project’s overall probability of success. We hypothesize that a qualitative capability shift is most likely to emerge when stepwise success ra¶52
  28. 4.1.1.6 DNA sequence design for Transcription Factor binding¶53
  29. Table 11¶56
  30. 4.2 Safeguards¶57
  31. 4.2.1 Trust-based access¶60
  32. 4.2.2 Model-level boundaries against harmful biological and cyber assistance¶65
  33. 4.2.2.1 Biological and Chemical boundaries¶66
  34. 4.2.2.2 Cyber boundaries¶70
  35. 4.2.3 Monitoring, Review, and Revocation¶74
  36. 4.2.4 Security and Access Controls¶79