thebench.bio #1 Ecosystem for the Bio-Tech Industry

Software Daily Summary

Software Daily Summary: What Happened on November 12, 2026

Latch Bio benchmark puts AI agents through structural-biology tasks

SecureBio also began tracking changes in models’ biological-safety scores over time.

  • Updated
  • Read from 2 articles, 2 publications
  • L S

Today in brief

Latch Bio tested AI agents on 58 structural-biology tasks involving pathogen targets. The leading standard configuration passed 43.7% of them.1 SecureBio began publishing repeated BioTIER tests, which found safety-score changes in both directions across models from four developers.2

Also worth knowing

  • L

    Latch Bio tested 20 models. Fable 5.2 with Claude Code led standard configurations, with a 95% confidence interval of 32.3% to 55.1%.1

  • S

    BioTIER contains 542 prompts across three biological-risk levels. SecureBio measured Gemini 3.1 Pro’s refusal score rising from 46% to 74%, while Grok 4.7’s fell from 88.5% to about 80%.2

Also published