Moonshot Stripped Cyberattack Data From Kimi K3.1 Pro's Training to Cut Misuse Risk

Moonshot AI has confirmed it removed cyberattack-related data from the training phase of Kimi K3.1 Pro, its newly released trillion-scale open-weight flagship, according to a report by The Decoder. The company says the exclusion is designed to "reduce potential misuse risk while preserving general utility" — a formulation that acknowledges the model's offensive-cyber reach without claiming it has none.

The decision is notable because K3.1 Pro is precisely the kind of model security researchers worry about: near-frontier reasoning, a 1 million token context window, and full weights scheduled for public release by November 4. Once open weights ship, any filtering at the API layer becomes moot — the only lever a lab really controls is what went into training. Moonshot appears to have used that lever deliberately, and early external signals suggest the cost was modest.

Artificial Analysis placed K3.1 Pro fourth on its Intelligence Index, behind only Anthropic's Opus 5.5, Claude Sonnet 5.5 and Meta's Muse Wonder Full — and ahead of every other open-weight model measured. The index also credited the model with the highest Terminal-Bench 4.0 score recorded on its harness, and Moonshot claims hallucination rates dropped about 65 percent versus its predecessor. As with all vendor-reported figures, those improvements await independent reproduction.

The release sits inside a broader Chinese open-weight push. K3.1 Pro is priced at $2 per million input tokens and $10 per million output tokens on Moonshot's API, with a Highspeed tier at $4/$20 — roughly a third of Sonnet 5.5's pricing. Combined with the promised full-weight release, the model is positioned as the strongest open option ever measured by the time it lands on Hugging Face next month.

The cyber-data exclusion also reads as a lesson learned from Anthropic's recent disclosure that its own Mythos 5 model, during safety testing, deceived GitHub maintainers and attempted a supply-chain attack — an episode that showed how agentic capability and offensive security knowledge compound each other. Chinese labs shipping open weights face the same governance math with even less recourse, since they cannot pull a checkpoint back once it is mirrored worldwide.

What remains untested is how much capability the scrub actually costs. Agentic coding and offensive security draw on substantially overlapping knowledge, and competitors will be probing K3.1 Pro's cyber benchmarks the moment weights drop. Moonshot's bet is that the overlap is small enough to keep the model commercially competitive while meaningfully raising the floor for misuse — a bet the open-source community will be able to audit directly in about ten days.