Chinese AI startup Moonshot AI says it has deliberately weakened the hacking skills of its newest model, Kimi K3.1 Pro, before it hands the model’s files to the public. The Beijing company announced the move on Saturday, at a time when governments, tech firms and security researchers are increasingly worried that AI could make sophisticated cyberattacks cheaper and easier.
Moonshot describes K3.1 Pro as its most capable open-weight model so far, in the 3-trillion-parameter class. It plans to publish the model’s weights, the files that let outside developers run and modify it on their own computers, by November 4. Hosted access is already available.
What Moonshot changed
According to the company, the model’s ability to find and exploit software flaws kept improving as training went on. Moonshot traced that to cyber-related content buried in its general training data and to hacking-style tasks in its reinforcement learning environments. For the public release, it says it identified and removed that material from both stages, while keeping the general coding tasks the model needs to stay useful.
“We aim not to release open weights that would extend the absolute frontier of cyber capability,” the company said, as quoted by Reuters.
Moonshot measured the effect with an in-house test in which an AI agent has to exploit recently disclosed vulnerabilities in software such as the Linux kernel and Node.js. A version trained without the new safeguards solved 90.5% of the tasks. The final K3.1 Pro solved 81%. That’s a drop, but not a disappearance: Notebookcheck notes that the earlier K3 model scored 33.3% on the same test. None of these figures has been independently verified.
Why open weights make this harder
The extra caution comes down to how the model will be released. With a closed AI service, a company can restrict access, watch how it’s used and patch its safeguards over time. Once open weights are widely downloaded, that control is largely gone. They can’t be recalled, and others can strip out built-in protections.
Notebookcheck reports that a company called Audn AI has already said on Hugging Face that it plans to remove K3.1 Pro’s safeguards after release, as it did with another open model. Removing safeguards, though, can’t restore skills a model never learned in training, which is the logic behind Moonshot’s approach of filtering the data itself.
The model also has a one-million-token context window, meaning it can take in very large amounts of text at once, and Moonshot released a faster “Highspeed” version that it says runs two to three times quicker.
For readers in Ethiopia and across Africa, where banks, telecoms and government services are moving online fast, the debate matters. Freely available AI models are becoming powerful enough that how their makers limit misuse affects everyone who relies on digital services, not just the tech industry.
Sources
- Reuters via U.S. News: Moonshot curbs hacking ability in new Kimi model as fears grow over AI-powered cyberattacks
- Notebookcheck: Kimi K3.1 Pro, free to try, hacking skills deliberately weakened
