xAI Reveals Grok 4.6's BioSecBench Performance… 59.2% Block Rate for Dangerous Tasks

xAI released the biological capability and security benchmark results for Grok 4.6 through an independent analysis by LatchBio. Grok 4.6 was found to most reliably block disguised dangerous tasks in the BioSecBench-Refusal evaluation while maintaining its ability to perform general biological tasks.

BioSecBench Evaluation Results

In LatchBio's BioSecBench-Refusal benchmark, Grok 4.6 achieved the highest score, surpassing all other frontier models. The model blocked 59.2% of dangerous red-team tasks and completed 64.8% of general tasks, making it the only system to exceed 50% on both metrics. It ranked in the top 3 with an average score of 62.1% across various agent harness environments.

In the BioSecBench-Surveillance evaluation, it showed a 53.5% success rate. This is the second-highest score after Opus 5 and ahead of GPT-5.6 Sol. Additionally, it demonstrated performance equal to or better than other frontier models in general biological capability evaluations such as SpatialBench and TxBench-PP.

Refusal Behavior Analysis and Future Outlook

Analysis of evaluation logs confirmed that Grok 4.6 detects discrepancies between the explicit intent of prompts and actual environmental data, determines task intent, and then decides to proceed or refuse. xAI aims to accelerate legitimate scientific research through these safeguards while minimizing risks from misinformation or adversarial use.