AISI's ReproBAIT benchmark tests whether coding agents can rebuild biological AI tools from scratch
charlieharris01 · x · 2026-10-27
The AI Security Institute released ReproBAIT (arXiv:2610.19810), an evaluation of whether coding agents can design, implement, and train published biological AI tools under restricted conditions.
- Agents run in an internet-free sandbox, receiving only training data, fixed time/compute budgets, and limited descriptions of the target tool
- Reproductions are scored on held-out data with tool-specific metrics, plus a memorisation-check methodology since open-source tools are likely in pretraining corpora
- Tested tools span increasing complexity: ProteinMPNN (inverse protein folding), RFdiffusion3 (all-atom generative diffusion), and a closed-source molecular design tool
- Findings: substantial reproduction capability for some tools among leading open/closed coding agents, but limited overall
The poster also quips that the UK government now has an official position that AlphaFold 3 and ESMFold are harder to reproduce than RFdiffusion3.
More from Safety
- AI safety researchers: a few lab staff full-time could make US-China AI deal verification feasible — ohlennart · 2026-10-28
- IAS advisory group demands OpenAI release withheld Yang-Mills and QCD proofs — rohanpaul_ai · 2026-10-28
- INT21: Agent swarms + 1,700 tok/s inference to outrun 27-second AI network attacks — bingxu_ · 2026-10-27
- Stop asking how smart your agent is — its permission blast radius matters more than model upgrades — Internal-Bug-8434 · 2026-10-27
- UK Home Office May Face Investigation Over Use of OpenAI Tools — wiredmagazine · 2026-10-27
- Luiza Jarovsky: governments' 3-day stockpile guidance reflects AI cyber risk — LuizaJarovsky · 2026-10-27