Moonshot filters cyber training from Kimi K3.1 Pro before open-weight release

Moonshot says targeted training-data and reinforcement-learning filters cut Kimi K3.1 Pro’s xdev-bench-hard pass@3 score from 90.5% to 81.0% ahead of a planned November 4 weights release.

2 min read

Code displayed across multiple monitors, illustrating software security testing and AI-assisted vulnerability research.

Code displayed across multiple monitors, illustrating software security testing and AI-assisted vulnerability research.

Photo by Markus Spiske on Unsplash (Unsplash License).

Moonshot AI says it deliberately reduced the offensive-cyber capability of Kimi K3.1 Pro before a planned open-weight release. The company announced the model on October 24 and says the hosted API is available now, while the full weights are due by November 4. Moonshot has not yet disclosed the license that will govern those weights, so “open-weight” should not be read as confirmation of an open-source license.

The intervention targeted two stages of training. Moonshot says it removed cyber-related material from the mid-training corpus and filtered cyber tasks from its reinforcement-learning environment mix, while retaining general-purpose coding work. On the company’s internal xdev-bench-hard evaluation, an unfiltered checkpoint reached 90.5% pass@3; the release candidate scored 81.0%.

That benchmark is more concrete than a generic safety claim. Its expert-curated tasks reproduce recently disclosed vulnerabilities in software including the Linux kernel, Node.js, JavaScriptCore, Exim and VirGL. The model receives vulnerable source code, a short bug hint and a binary, then must achieve arbitrary code execution in a separate sandbox and retrieve a secret string. Moonshot reports pass@3 on a random subset of runs, with no hard time or output-token limits.

The result also has clear limits. It is a publisher-run, in-house benchmark and has not been independently verified. One lower score does not establish that the model is safe, and Moonshot says K3.1 Pro still substantially outperforms its predecessor on the same cyber evaluation. Once downloadable weights are distributed, the developer also has less control over access and later safeguards than it does with a hosted service.

K3.1 Pro is a thinking-only, natively multimodal 3T-class model with a one-million-token context window. Moonshot also offers a Highspeed variant that it says preserves model intelligence at roughly two to three times the speed.

Sources: Moonshot AI technical blog · Kimi API guide · Kimi model list · Reuters

benchmarksAI securityMoonshot AIKimicybersecurityopen weightsKimi K3.1 Pro