Moonshot Filters Cyber Data From Kimi K3.1 Pro and Will Release Open Weights by Nov. 4

Moonshot AI said on Saturday, October 24, that its new Kimi K3.1 Pro model scores 81.0% on an internal test of finding and exploiting software vulnerabilities, against 90.5% for a similar version built without cyber-risk measures. The Beijing company plans to publish the weights by November 4, according to a Reuters report.

Moonshot said it removed cyber-related data from the training corpus and from the reinforcement-learning environments, while keeping general coding skills. "We aim not to release open weights that would extend the absolute frontier of cyber capability," the company said.

Nobody outside Moonshot has run the benchmark. The Reuters report does not name the test, count its tasks or say who built it, and describes only a controlled environment.

What the 81.0% does and does not show

The three figures below are all Moonshot's. The Kimi K3 score comes from Crypto Briefing's account of the announcement.

Model	Moonshot internal exploit test	Status
Kimi K3	33.3%	Weights published in July
Kimi K3.1 Pro	81.0%	Weights due by Nov. 4
Similar version without cyber measures	90.5%	Not released

The filtering cost 9.5 points, about a tenth of the unfiltered score. The curbed model still scores 2.4 times what Kimi K3 scored on the same test. On Moonshot's own numbers, the curb lowers a ceiling that the company raised from 33.3% to 90.5% in the same release.

What outside testers found on Kimi K3

The only independent figures are for the earlier model. A joint assessment by the UK AI Security Institute and CAISI, the US government's NIST-based evaluation centre, found that Kimi K3's safeguards did not stop it attempting exploit development or offensive operations, according to Forkast, in a report dated July 25.

On ExploitBench, which uses 41 V8 engine vulnerabilities from after 2023, Kimi K3 scored 32%, against 24% for GLM-5.2 and an average of 76.2% for US frontier models. It succeeded zero times in 41 on arbitrary code execution tasks. On a 32-step simulated network attack it reached step 17 on average, against 28.5 for the US models.

Those numbers sit close to Moonshot's 33.3% for the same model, but they come from a different test. They cannot be used to convert the new 81.0% into an outside-tested score.

China's draft export rules

China proposed export controls last week that would ban exports of AI-driven malware generation and mutation technology and restrict AI-assisted vulnerability-discovery tools. The draft remains open for public comment, Reuters reported. The report does not say whether published model weights would fall under it.

Moonshot argues that open weights can improve security by letting organisations inspect and secure their own software, including older systems. Reuters gave the counterpoint: once files are widely distributed, a developer has less ability to restrict access, monitor use or update safeguards than with a hosted service. The same concern drove OpenAI's decision to move frontier training offline.

Specifications and what is missing

Crypto Briefing reports a mixture-of-experts design with 2.8 trillion total parameters, 104 billion active, and a 1 million token context window. Reuters calls it a "3-trillion-parameter-class" model and says a K3.1 Pro Highspeed variant runs two to three times faster.

Crypto Briefing also reports 65.7% on Terminal-Bench 4.0, which it describes as comparable to Claude Opus 5.5. That is a vendor-supplied figure with no independent run behind it.

Neither report states a licence for K3.1 Pro. Kimi K3 shipped under a Modified-MIT licence, per Forkast. As of Saturday, Moonshot's Hugging Face organisation page listed Kimi-K3, at 2.8 trillion parameters, as its newest model and did not list K3.1 Pro. Compare how Qwen3.8-27B shipped weights and licence together.

Moonshot said it will publish the weights by November 4. Outside testers can check the cyber claims only after that.