Moonshot deliberately limits Kimi K3.1 cyber capabilities: 74.7 vs 100 on Exploitbench
bookwormengr · x · 2026-10-24
Moonshot intentionally reduced cyber capabilities in the open-weights Kimi K3.1: 74.7 on Exploitbench (vs 100 for GPT-6 Astra) and 20.7 on ExploitGym (vs 42.7). This was done by filtering datasets rather than RL on such tasks, since these capabilities were 'emerging' on their own.
The official doc lays out open-weight principles:
- Never release open weights that extend the absolute frontier of offensive cyber capability; K3.1 Pro is deliberately below what Moonshot's pipeline can produce
- Progressively raise cyber capability in open weights; non-discriminatory access benefits defenders auditing code, including legacy systems
- Gradual diffusion of capability beats concentrating it in a few proprietary systems for overall software security
The author calls this a better approach than shipping highly capable models with easily-removed safeguards.
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
- 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
- Hand-crafted regex pretokenizers quietly do heavy lifting in every frontier LLM — YouJiacheng · 2026-10-24