Can AI coding agents rebuild biological AI tools from a paper?

Impact 8 AISI Blog

Summary

  • The UK AI Safety Institute (AISI) and SecureBio developed the ReproBAIT evaluation to test if AI coding agents can reproduce biological AI tools (BAITs) from limited public information.
  • Results show that current frontier coding agents can implement and train functional replicas of less complex BAITs using only a paper and training data within hours and for under $20.
  • The most complex tool, RFdiffusion3, was not successfully reproduced within the given single-GPU time and compute budget.
  • Withholding code or limiting published information does not provide effective protection against misuse, suggesting that data governance may be necessary for managing biological AI risks.

Why it matters

Demonstrates that frontier AI coding agents can reverse-engineer biological AI tools from published literature and training data alone, challenging current assumptions that withholding code is sufficient to prevent dual-use misuse.

Who should care