Grok 4.6 Benchmarked: Strong Biosecurity Without Capability Loss

kenbwork · x · 2026-09-01

LatchBio released a report benchmarking Grok 4.6's biological capabilities and security. The model performs well at rejecting dangerous red-team queries while answering legitimate research inquiries. Safeguards do not degrade biological capabilities, with Grok 4.6 placing 4th on the overall leaderboard. Its refusal behavior is driven by intelligence rather than input classifiers.

Related event: Grok 4.6 passes LatchBio biosecurity benchmark, refusing dangerous queries without hurting research(5 posts)→

Original post →

More from Models

• Mocked benchmark shows DeepSeek's document chart grounding badly misaligned — burkov · 2026-10-18
• Dev complains Claude's habit of saying "not done" is his biggest pain point — jasonkneen · 2026-10-18
• Open-weight TYGER speech decision model makes 80ms turn-taking calls for full-duplex voice agents — solyarisoftware · 2026-10-18
• Paying Google AI subscriber blocked: custom-domain Workspace accounts can't access AI Studio — ChrisRackauckas · 2026-10-18
• Running Ling 3.1 Flash (560B MoE, 25B Active) Across 4 DGX Sparks via llama.cpp RPC — BerryBlush_Rogers691 · 2026-10-18
• Users press OpenAI on long-delayed Adult Mode a year after it was promised — ___Patrice___ · 2026-10-18

Models channel →