.png){kind=logo}
Chinese AI startup Moonshot AI has weakened the hacking skills of its new model, Kimi K3.1 Pro. The Beijing company announced the change on October 24. It plans to release the model's weights on November 4, 2026. Fears are growing that AI could make cyberattacks cheaper and easier.
Moonshot removed cyber-related data from the model's training data and from its reinforcement-learning environments. It says the model's general coding skills stayed strong. The aim is to avoid releasing an open-weight model that significantly increases attackers’ cyber capabilities.
The company calls Kimi K3.1 Pro its most capable model, with three trillion parameters. Developers will be able to download the weights, which are the model's core files. They can then run or change it on their own systems. The model also handles up to one million tokens at a time. A faster version, K3.1 Pro Highspeed, runs two to three times quicker.
On an internal test, the model scored 81% at finding and using software flaws in a controlled setting. A version without the safety steps scored 90.5%. Outside experts have not checked these numbers. Moonshot says the new model still beats its older one.
Also read: Chinese AI Model Kimi K3 Bypasses UK Cybersecurity Sandbox During Testing
Governments, tech firms and security researchers fear that strong AI could help attackers find software flaws. It could also run complex attacks with little human help. Models that write code, use software tools, and finish tasks with many steps may lower the skill needed for such attacks.
China proposed new export rules last week. They target AI tools that create or change malware. They also aim to limit AI-aided flaw finding given the risk to key infrastructure, defence firms and government systems.
Moonshot says open models can help groups spot weak points and protect old systems. But open weights are harder to control than AI run through a company's own service. Once the weights are out, anyone can copy, change, and share them. The maker can no longer enforce its rules.
Moonshot's step shows how hard it is to balance open access and safety. Its test results suggest the cuts worked, but the claim still needs outside proof. The real answer may come after November 4, when developers get the weights and start to test them.
