LatchBio Evaluates Grok 4.6 on Biosecurity and Adversarial Biological Tasks

Key Info

LatchBio evaluated Grok 4.6 on biosecurity monitoring and adversarial biological tasks, reporting that the model correctly detects and refuses dangerous queries—even when maliciously obfuscated—while still answering legitimate scientific questions.

Highlights

• Grok 4.6 successfully identifies and blocks dangerous biological requests, including obfuscated adversarial attempts.
• Beneficial scientific queries remain accessible, showing a balanced safety approach.
• Full results and discussion are available in LatchBio's blog post.