Moonshot reveals it stripped cyber-offense training data from open-weight Kimi K3.1 Pro
teortaxesTex · x · 2026-10-24
Kimi K3.1 Pro's report discloses that cyber content was naturally embedded in the general mid-training corpus and RL task mix, steadily boosting vulnerability-research capability. Internal experiments showed downstream cyber capability would rise substantially without data removal, so for the open-weight release Moonship identified and removed cyber-related data from both training stages while keeping general-purpose coding tasks, validating with in-house benchmark xdev-bench-hard (reproducing real CVEs in Linux kernel, Node.js, JavaScriptCore, Exim, VirGL; pass@3). Commentators note this is an official admission of targeted cyber-capability nerfing, and debate whether training only on defense inevitably improves offense too.
More from Models
- Community claims Kimi k3.1 is the most 'benchmaxxed' model ever, alleging distillation — imjustnewatai · 2026-10-24
- Leaker: Grok 4.8 arriving next week — mark_k · 2026-10-24
- Kimi K3.1 Costs ~3x DeepSeek Flash Per Task but Nearly Doubles Terminal-Bench 4 Score — bookwormengr · 2026-10-24
- Labs are far ahead of what they release — and the gap may be widening, speculates X user — MoonL88537 · 2026-10-24
- API Pricing Cut from $15 to $10 per Million Tokens — cedric_chee · 2026-10-24
- MiniMax's 430B model beats 3T-scale rivals on AA knowledge benchmark — teortaxesTex · 2026-10-24