Moonshot unveils Kimi K3.1 Pro, deliberately dialing down its hacking skills

What's going on

Beijing-based AI startup Moonshot AI unveiled its new model, Kimi K3.1 Pro, on October 24. Moonshot calls it its most capable open-weight model in the three-trillion-parameter class. "Open weight" means the weights, the values the model learned during training, are published as files, so other developers can run and modify the model on their own systems. Moonshot says it will make these weights available by November 4 at the latest. For now, they can't be downloaded yet.

The announcement stands out because Moonshot explicitly says it deliberately limited the model's capabilities for cyberattacks. The background is growing concern that increasingly autonomous AI could make hacking much easier.

How the hacking skills were held back

According to the company, cyber-related data was removed from both the training corpus (the body of text the model learns from) and the reinforcement learning environments. In reinforcement learning, the model practices on simulated tasks and is rewarded for good results. General programming abilities are said to be preserved.

Moonshot states the goal like this:

We don't want to release open weights that push the absolute frontier of cyber capabilities.

Cyber benchmark numbers

On its own internal cyber benchmark, Moonshot measures how well the model finds and exploits software vulnerabilities in a controlled environment.

Moonshot says the comparison shows that filtering out cyber data lowers offensive performance. The model still outperforms its predecessor, though. Important: the results haven't been independently verified. A gap of a little over 9 percentage points is also modest, so the model remains clearly capable of such tasks.

Key technical details

K3.1 Pro has a context window of one million tokens. That's the amount of text the model can process at once. There's also K3.1 Pro Highspeed, which Moonshot says runs two to three times as fast as the standard model. The source gives no further details, such as the license or benchmarks outside the cyber domain.

Regulatory background

Governments, tech companies and security researchers increasingly worry that AI is making hacking easier and cheaper. Models that write code, operate software tools and carry out long multi-step tasks can help find vulnerabilities and automate parts of attacks.

Last week, China proposed export controls that would ban exports of AI-powered technology for generating and mutating malware and restrict AI-powered tools for vulnerability discovery. The draft is still open for public comment. It cites the risk of use against critical infrastructure, the defense industry and core government systems, as well as bypassing security safeguards.

The case for open models

Moonshot argues that open models can improve security because more organizations can test and harden their own software, including older systems that get little attention. The counterargument: once the files are out, the developer can hardly restrict access, monitor usage or update safeguards, unlike with a hosted service.

Our take

It's notable that the maker of an open model publicly throttles cyber capabilities as a design choice and backs it with numbers. It also fits the intensifying debate in China. How robust the measure is remains an open question: anyone with the weights can continue training the model and possibly restore those capabilities; the source says nothing on this. The benchmark figures also come from the maker itself. Real-world performance can only be independently tested once the weights are released.