Grok shares a biological-safety evaluation of version 4.6
The September report describes tests that weigh routine research usefulness alongside refusals of hazardous tasks.
THE STORY IN THREE POINTS
- The official evaluation report appeared on September 1.
- It covers routine work and hazardous-task refusal.
- The reported results are specific to Grok 4.6 and the tested setups.
The announcement
The Grok team published a September 1 account of LatchBio’s evaluation of Grok 4.6. It describes separate benchmarks for distinguishing hazardous biological tasks from routine research and for surveillance workflows. The report presents both refusal and ordinary-task completion figures and says multiple agent harnesses were tested. This brief attributes those findings to the official report; it does not treat the results as a comprehensive safety certification.
Editorial context
A refusal score is difficult to interpret without knowing which useful tasks were also blocked. Evaluating both sides makes the tradeoff more visible. The broader editorial lesson applies to specialized assistants: a system must be judged against the actual range of intended work, including situations where the right response is to request more context or decline an action.
Editorial context: what to watch
Future evaluations could publish fuller protocols, version identifiers and independently reviewable results without exposing dangerous details. Tests should also be refreshed as tools and model settings change. The existence of a benchmark is a starting point for comparison; confidence in a deployment depends on how closely the tested conditions match the workflow in which it will be used.
Sources: Official report on LatchBio’s Grok 4.6 evaluation, September 1 ↗
