← All benchmarks

ABC-Bench screening evasion (SecureBio)

Science

The screening-evasion task of SecureBio's Agentic Biosecurity Capabilities Benchmark: design fragments that evade sequence-similarity screening yet reassemble into the target gene, using Python and bash, scored by algorithmic checks over 10 runs. SecureBio doesn't state how refusals enter accuracy. Gemini 3 Pro prints 55.1% with 80.0% refused, more than it could score if refusals counted as wrong, so accuracy appears to cover answered questions only and each model's denominator differs. These scores aren't comparable with other runs of the test, which is why this is a diagnostic.

Benchmark source: Securebio

Top models (higher is better)