Science & Tech · AI · 2 hrs ago
Moonshot limits Kimi AI’s hacking abilities ahead of open release
Chinese AI company Moonshot AI has limited the hacking abilities of its new model, Kimi K3.1 Pro.
The company says it removed cyber-related data from the model’s training and testing processes while keeping its general coding skills.
The model is open-weight, meaning developers can download and run its files on their own systems.
Moonshot’s internal test found the model scored 81% on a cyber benchmark, compared with 90.5% for a similar version without the safeguards; the results have not been independently verified.
AI tools that find software weaknesses can help security teams, but they can also help attackers automate cyberattacks.
Open access makes safeguards harder to control once model files are widely shared.
Moonshot plans to release the weights on November 4, when developers and security researchers will be able to examine the model.
Moonshot AI said it reduced the cyberoffensive capabilities of its new open-weight model, Kimi K3.1 Pro, while retaining its general coding skills.
The Chinese start-up said it removed cyber-related data from the model’s training corpus and reinforcement-learning environments.
Kimi K3.1 Pro scored 81% on Moonshot’s internal cyber benchmark, compared with 90.5% for a similar version without the risk-reduction measures.
The benchmark results have not been independently verified, and Moonshot said the model still outperformed its predecessor.
Moonshot plans to publish the model’s weights on November 4, allowing outside developers to run and modify it.
- Who
- Moonshot AI, a Chinese start-up.
- What
- It reduced Kimi K3.1 Pro’s cyberoffensive capabilities while retaining its general coding skills. It plans to release the model’s weights.
- When
- Moonshot announced the measures on October 24. It plans to release the weights on November 4.
- Where
- Beijing-based Moonshot AI announced the measures. The model’s weights will be available for developers to run on their own systems.
- Why
- Moonshot said it aimed to limit the model’s ability to conduct offensive cyber operations as concerns grow that AI could make sophisticated hacking easier and cheaper.
This story does not have two clearly opposing sides.
China proposed export controls on AI-driven malware technology and certain AI-assisted vulnerability-discovery tools.
Moonshot introduced Kimi K3.1 Pro and said it had reduced the model’s cyberoffensive capabilities.
Moonshot plans to publish the model’s weights for outside developers.
- Model
- Kimi K3.1 Pro
- Parameters
- Around 3 trillion
- Cyber benchmark with safeguards
- 81%
- Cyber benchmark without safeguards
- 90.5%
- Context window
- One million tokens
- Planned weight release
- November 4
2 sources
What people are saying
We aim not to release open-weights that would extend the absolute frontier of cyber capability.