Moonshot admits deliberately removing cyber training data from open-weight Kimi K3.1 Pro

In the Kimi K3.1 Pro technical report, Moonshot disclosed that as the model grew more capable, cybersecurity content naturally crept into general mid-training corpora and RL environments, steadily boosting vulnerability research ability through post-training. The team therefore proactively scrubbed the relevant data and intentionally suppressed the open release's offensive/defensive cyber capabilities, stating that downstream cyber ability would otherwise rise substantially.

Confirmed

• The reduction is disclosed in the technical report; multiple authors including @teortaxesTex, @scaling01, and @bookwormengr relayed the same fact
• Internal benchmark xdev-bench-hard (pass@3): the un-reduced checkpoint scored 90.5, dropping to 81.0 after data removal
• ExploitBench score of 74.7 (GPT-6 Astra: 100); ExploitGym score of 20.7 (GPT-6 Astra: 42+)
• @scaling01 notes that even with the data retained, internal cyber evaluations would still trail GPT-5.6-Sol

Why it matters

• A rare case of a vendor explicitly admitting capability-specific weakening in a technical report, with direct implications for the cybersecurity research value of open-weight models
• The quantitative data makes the "safety alignment vs capability ceiling" tradeoff assessable: nearly 10 points separate the un-reduced checkpoint from the released version
• Users relying on open models for vulnerability research or security testing should expect K3.1 Pro's cyber performance at its suppressed level

2026-10-24 ~ 2026-10-24 · 10 related posts

• [source] Moonshot reveals it stripped cyber-offense training data from open-weight Kimi K3.1 Pro — teortaxesTex · 2026-10-24
• [source] Kimi K3.1 Pro ships nerfed on cyber: internal bench drops from 90.5 to 81.0 after data removal — teortaxesTex · 2026-10-24
• Moonshot lays out open-weight principles: never ship weights that extend the cyber frontier — teortaxesTex · 2026-10-24
• [source] Moonshot says K3.1 Pro's offensive cyber capability is deliberately capped below frontier — dejavucoder · 2026-10-24
• Moonshot deliberately curbed Kimi K3.1's cyber capabilities: 74.7 on ExploitBench vs GPT-6 Astra's 100 — bookwormengr · 2026-10-24
• Moonshot deliberately limits Kimi K3.1 cyber capabilities: 74.7 vs 100 on Exploitbench — bookwormengr · 2026-10-24
• Moonshot removed cyber data from Kimi training to curb downstream attack capability — AashaySachdeva · 2026-10-24
• Moonshot published K3.1 cyber capability scores and stripped cyber data from its open-weight release — cedric_chee · 2026-10-24
• Moonshot removed cyber-related training data from open-weight Kimi K3.1 Pro — nrehiew_ · 2026-10-24

1 near-duplicate retellings: scaling01