Safety

Grok 4.6 Shines in Biosecurity Benchmarks

Grok 4.6 outperformed other models on BioSecBench-Refusal, scoring above 50% on both refusal and routine tasks.

Image: xAI

Grok 4.6 demonstrated strong biosecurity capabilities in independent testing.

LatchBio's analysis found Grok 4.6 to be the strongest model at refusing disguised hazardous tasks while still completing routine biological work.

The model scored above 50% on both measures, a feat no other system achieved.

"Grok 4.6 is the only model tested that scored above 50% on both measures," said LatchBio.

The company emphasized its ongoing efforts to ensure safe deployment of frontier models.

Source: xai

Grok 4.6 scored above 50% on both refusal and routine tasks in BioSecBench-Refusal.
Grok 4.6 is the only model tested that scored above 50% on both measures.
LatchBio evaluated Grok 4.6 on BioSecBench-Refusal and BioSecBench-Surveillance.
Grok 4.6 averages a success rate of 53.5% on BioSecBench-Surveillance.
Grok 4.6 outperformed other models on BioSecBench-Refusal.
Grok 4.6 holds the top three spots across different harnesses, averaging 62.1%.
Grok 4.6 refused 59.2% of red-team tasks and completed 64.8% of routine ones.

WRITTEN BY

AI Safety, Alignment & Policy

Nadia follows AI safety, alignment, regulation, and the policy debates shaping the field.

Related articles