Chinese artificial intelligence company Moonshot AI has deliberately reduced the cyber-offensive capability of its latest Kimi model before releasing it publicly, reflecting growing concern that powerful open AI systems could make sophisticated hacking cheaper and easier.
The Beijing-based company said it removed cyber-related data from the training process for Kimi K3.1 Pro while trying to preserve its broader software-development abilities.
Moonshot plans to release the model’s weights on November 4.
That would allow developers to download, run and modify the AI on their own infrastructure.
FCRF’s Chief AI Officer Certification Returns on November 14 to Prepare Future AI Leaders
Cyber Training Data Was Removed
Moonshot said cyber-related information was removed from both the model’s training corpus and its reinforcement-learning environments.
Reinforcement learning is a stage where an AI system improves by attempting tasks and receiving feedback on its performance.
If those training environments include vulnerability discovery and exploitation exercises, the model can become increasingly capable at offensive cybersecurity tasks.
Moonshot decided to limit that exposure.
The company said its goal was not to publish open weights that pushed beyond the existing frontier of offensive cyber capability.
The approach differs from simply placing safety filters around an AI chatbot.
Once an open-weight model is downloaded, users may be able to modify or remove external restrictions.
By changing the training process itself, Moonshot is attempting to reduce risky capability inside the model before the files are distributed.
Internal Hacking Score Fell From 90.5% to 81%
Moonshot says the safety intervention produced a measurable reduction in cyber performance.
On an internal benchmark that tested whether the model could identify and exploit software vulnerabilities in a controlled environment, the released Kimi K3.1 Pro scored 81%.
A similar version developed without the cyber-risk reduction measures scored 90.5%.
That represents a 9.5-point drop.
However, the figures should be treated carefully.
The benchmark and results come from Moonshot itself and have not been independently verified.
An 81% score also does not mean the model has lost its ability to assist with security research or exploitation.
Moonshot says K3.1 Pro still performs better on cyber tasks than its previous model.
The company is therefore limiting the capability rather than removing it.
Open Models Create a Different Cybersecurity Problem
The decision highlights a growing tension around open-weight AI.
Open models can benefit researchers and companies because they can be inspected, customised and operated without sending sensitive information to an external provider.
Moonshot argues that such systems can help organisations analyse and secure old software that might otherwise receive little attention.
But open weights also reduce the developer’s control.
A hosted AI provider can monitor use, restrict accounts or update safeguards after detecting abuse.
Once model weights are downloaded and distributed widely, those controls become much harder to enforce.
Developers may modify the model, remove restrictions or connect it to automated software tools.
That becomes more concerning as models improve at writing code, operating computers and completing long sequences of tasks without constant human supervision.
Cybersecurity researchers have already documented hackers using AI agents to scan large numbers of websites, identify weaknesses and attempt exploitation with limited human involvement.
The420.in recently reported a campaign in which one hacker used autonomous AI agents to attack more than 1,200 websites and breach over 100.
Governments Are Starting to React
Moonshot’s move also comes as governments begin paying greater attention to offensive AI capabilities.
China recently proposed export controls that would restrict AI-powered malware generation and some AI-assisted vulnerability-discovery technologies.
The draft measures specifically cite the risk that advanced cyber tools could be used against critical infrastructure, government systems and defence-related organisations.
Elsewhere, technology companies and security researchers are debating how much cyber capability should be included in publicly available AI models.
There is no simple dividing line.
The same AI system that helps a security researcher discover a dangerous vulnerability could potentially help a criminal find the same weakness.
Limiting offensive capability too aggressively could make defensive research harder.
Allowing unrestricted capabilities could lower the expertise needed to conduct sophisticated attacks.
Moonshot’s approach represents one possible compromise: preserve general coding and defensive usefulness while deliberately reducing performance on exploitation tasks.
Kimi K3.1 Pro also includes a one-million-token context window, allowing it to process extremely large quantities of text and code in a single session.
The company has released a Highspeed version that it says operates two to three times faster than the standard model.
The more important test will come after the open weights are released.
Researchers will then be able to independently examine whether the cyber restrictions remain effective once the model is running outside Moonshot’s own systems.
What this means for you:
AI developers are starting to treat hacking capability as something that may need to be deliberately limited before powerful models are released. For open models, that decision matters more because safeguards become far harder to enforce once the underlying model files are publicly distributed.
Follow for daily updates on cybercrime, corporate fraud, DFIR, hacking, investigations, and digital forensics
