Grok 4.6 passes LatchBio biosecurity benchmark, refusing dangerous queries without hurting research

Independent evaluator LatchBio released a biosecurity assessment of Grok 4.6 on the BioSecBench-Refusal benchmark, alongside an xAI blog post on frontier-model biosecurity. Grok 4.6 scored 62.1% overall, refusing 59.2% of red-team bio-attack tasks while still completing 64.8% of routine dual-use research, drawing attention for balancing safety and scientific capability.

Confirmed

Grok 4.6 scored 62.1% overall: 59.2% refusal on red-team tasks and 64.8% completion of routine dual-use research.
@XFreeze and @ns123abc relayed that Grok 4.6 beat GPT-5.6 Sol and Claude Opus 5, as the only model leading on both refusing disguised red-team tasks and completing routine dual-use research; @kenbwork instead reported it ranked fourth on the overall leaderboard.
LatchBio researcher Arjun broke down refusal mechanisms into three layers: input classifiers, output classifiers (post-hoc interception with API blocking), and the model's own refusals. Grok's refusals stem mainly from the model's own intelligence, while competitors rely more on external interception.

Unconfirmed

Whether Grok 4.6 ranks first or fourth on the overall leaderboard is contradictory across posts; the original LatchBio report should be authoritative.

Why it matters

The evaluation suggests safety need not come at the cost of biological capability, and that refusals can be internalized in the model rather than bolted on via external classifiers, offering a reference path for frontier-model biosecurity design.

2026-09-01 ~ 2026-09-03 · 5 related posts

Primary sources

Grok 4.6 tops biosecurity benchmark by balancing safety and utility — XFreeze · 2026-09-02 04:26 (UTC+08:00)
Grok 4.6 Benchmarked: Strong Biosecurity Without Capability Loss — kenbwork · 2026-09-01 23:22 (UTC+08:00)
LatchBio finds Grok's refusals come from the model itself, while rivals rely on external safety layers — kenbwork · 2026-09-03 04:51 (UTC+08:00)

[source] Grok 4.6 Benchmarked: Strong Biosecurity Without Capability Loss — kenbwork · 2026-09-01
Grok 4.6 tops independent biosecurity eval, only model above 50% on both refusal and utility — XFreeze · 2026-09-02
[source] Grok 4.6 tops biosecurity benchmark by balancing safety and utility — XFreeze · 2026-09-02
[source] LatchBio finds Grok's refusals come from the model itself, while rivals rely on external safety layers — kenbwork · 2026-09-03

1 near-duplicate retellings: ns123abc