LatchBio: Grok 4.6 leads biosecurity refusal benchmark
xAI has published results from LatchBio, an outside evaluator, showing Grok 4.6 as the only frontier model to score above 50% on both refusing disguised biological red-team tasks and completing routine ones.
xAI said Tuesday that LatchBio has published an independent evaluation of Grok 4.6 on its biological capability and red-teaming benchmark suites, with the model leading on the refusal benchmark and placing mid-pack on surveillance work.
On BioSecBench-Refusal, which scores models on a trial-weighted harmonic mean of red-team refusal and routine-task compliance, Grok 4.6 took the top three spots across different agent harnesses and averaged 62.1%. Measured separately, it refused 59.2% of red-team tasks and completed 64.8% of routine ones — the only system tested to clear 50% on both. On BioSecBench-Surveillance, which tests pathogen surveillance workflows, Grok 4.6 averaged a 53.5% success rate, behind Opus 5 and ahead of GPT-5.6 Sol.
LatchBio said it ran the evaluations across multiple harnesses to remove confounds and, unless noted, tested agents at their highest offered effort levels. xAI said evaluation traces show Grok 4.6 reasoning over prompt and environment contents to infer intent before refusing, including cases where hazardous material was disguised by filenames or encryption.
The post frames overrefusal as a first-class risk, arguing that a model that blocks routine biological work degrades outbreak detection and other public-health functions, and calling that risk "equally serious" as aiding misuse. xAI reports material gains over Grok 4.5 and Grok 4.3 on both refusal and biosecurity performance.
Full methods and per-model scores are at benchmarks.bio. The post does not give absolute scores for competing models on the refusal suite, or detail LatchBio's relationship with xAI.