Biosecurity at the frontier

What changed

Source summary

Independent testing by LatchBio found Grok 4.6 was xAI's strongest model yet at refusing disguised or hazardous biological task requests while still completing legitimate scientific work, following xAI's rollout of a layered safety approach combining refusal training, inference-time safeguards, behavioural controls and post-deployment monitoring.

Why it matters to your work

Practical interpretation

Not something most people will action directly, but reassuring if you work in health, biotech or research and are considering Grok: independent testing shows it is now better at refusing misuse of biological information while still helping with genuine scientific questions.

When this could help

Relevant for updating AI risk assessments and evaluating frontier model biosecurity posture, particularly if your organisation uses or is considering Grok for biology-related work.

An independent evaluation by LatchBio found Grok 4.6 is the only tested frontier model that both refuses over 50% of red-team biosecurity tasks and completes over 50% of routine biology tasks (59.2% refusal, 64.8% routine completion). It uses layered safeguards including refusal training, inference-time blocking, behavioural controls, and post-deployment monitoring.

Build it into your routine

SCALA suggested workflow

1. Biosecurity evaluation tracking

   When evaluating frontier models for biology-adjacent work, check their BioSecBench scores and refusal rates alongside capability benchmarks.

What to check

• Pricing, specific availability dates, and plan tiers for Grok 4.6 are not stated in this article — it focuses on safety evaluation, not product access. Source ↗
• xAI acknowledges that overrefusal is a risk it considers as serious as aiding malicious use. Source ↗