xAI quietly revised Grok 4.6 model card, cutting harmful cyber compliance from 16.7% to 6.9%
Miles_Brundage · x · 2026-08-21
Safety watchdog TheMidasProj reports that xAI quietly revised the published Grok 4.6 model card on Aug 17 — the second post-release revision after Grok 4.5. This update includes a changelog, but it is incomplete and offers no explanations for the corrections.
Among the four corrected evals: HackerBench harmful/dual-use cyber compliance fell from 16.7% to 6.9% (originally more than double Grok 4.5's 7.8%); self-harm compliance fell from 3.7% to 0.84%, still worse than Grok 4.5's 0.50%. Multiple safety evals moved in Grok 4.6's favor, though self-harm and MASK dishonesty still trail the previous model.
Related event: xAI Quietly Revises Grok 4.6 Model Card After Release(2 posts)→
More from Models
- Zhipu's SAO: single-rollout async RL trains stably for 1,000 steps, beats GRPO — teortaxesTex · 2026-08-21
- Gemini's Safety Filters Too Strict? Rejects Kissing and Roadside Photos — Dry-Sympathy-3182 · 2026-08-21
- 0.63M Parameter Verifier Matches 7B Models in Specific Tasks — jm_alexia · 2026-08-21
- Pangram v4 Model Claims to Remove AI Watermarks and Mimic Human Writing — Scobleizer · 2026-08-21
- Brundage: something funky going on with ChatGPT inference, likely testing — Miles_Brundage · 2026-08-21
- Does Claude perform better in 'claudish'? Researchers call for empirical measures — voooooogel · 2026-08-21