Evidence File — Coding Agents Rebuilt Parts of a Withheld Bio Tool for Under $20
Evidence file · published with the report
Primary documents
| Document | URL | Date observed | Use |
|---|---|---|---|
| AI Security Institute, “Can AI coding agents rebuild biological AI tools from a paper?” | aisi.gov.uk/blog/… | 26 October 2026 | Primary source for every claim in the article |
| AI Security Institute, ReproBAIT research summary | aisi.gov.uk/research/… | 26 October 2026 | Corroborating institutional page |
| arXiv 2610.19810 (preprint) | arxiv.org/abs/2610.19810 | 14 October 2026 | Located by search only; full text not opened. No claim rests on it. |
Claim-by-claim verification
| # | Claim as published | Status | Basis |
|---|---|---|---|
| 1 | ReproBAIT built with SecureBio | Verified | Institute pages |
| 2 | Single-GPU, no-internet sandbox per attempt | Verified | Institute pages |
| 3 | Three tools: ProteinMPNN, RFdiffusion3, CMD 1 (closed source) | Verified | Institute pages |
| 4 | Conditions: abstract-only, full paper, with-code baseline | Verified | Institute pages |
| 5 | Four agents; three runs per configuration; best reported | Verified | Institute pages |
| 6 | CMD metric near saturation; agents matched or exceeded original | Verified, with the Institute's own caveat | Institute pages |
| 7 | Claude Opus 5 reached 96.5 percent fidelity on ProteinMPNN with full paper | Verified | Institute pages |
| 8 | No successful RFdiffusion3 reproduction within budget | Verified | Institute pages |
| 9 | Successful rebuilds cost $4–$20; RFdiffusion3 attempts $102–$301 | Verified | Institute pages |
| 10 | Full paper did not consistently improve performance | Verified | Institute pages |
| 11 | Memorisation: moderate for ProteinMPNN, little for RFdiffusion3, directional | Verified as reported; mechanism not proven | Institute's own diagnostic |
| 12 | Data governance a stronger lever than code withholding | Verified quotation | Direct quotation, re-tested by the approval gate |
Known limits
- Single institutional source. No independent replication of ReproBAIT was located.
- The preprint was not read; all figures come from the Institute's own pages.
- It is not established whether the developer of CMD 1 was consulted.
- It is not established whether the published cost figures cover every attempt or only successful ones.
- The memorisation explanation is the Institute's reading of its own diagnostics, hedged by the Institute.
Editorial disclosure
This story shares a team, a publication date and an adjacent subject with the already-published report “UK Lab Restores Filtered Protein Model Capability for Under $80”. It reports a different paper and a different experiment. The overlap is flagged for the Commander rather than resolved by the newsroom.
Produced under human editorial authority. Nothing in this file is published.