Moonshot limits the capabilities of its most advanced AI model to prevent hacker attacks

Chinese startup Moonshot AI announced that it had reduced some of the capabilities of its newest artificial intelligence model, Kimi K3.1 Pro. The decision came amid growing concerns about the use of AI to make hacker attacks more sophisticated and cheaper.

Based in Beijing, the company describes the tool as its most advanced open model to date. With this setup, it allows external developers to run and modify the AI on their own systems.

Model did not lose general capabilities

• To limit misuse of the tool, Moonshot removed data related to cyberattacks from the training corpus and from the reinforcement learning environments used in the model's development.
• According to the company, this decision prevents improper uses while preserving general programming skills.
• The effects of this filtering were measured by an internal cybersecurity benchmark: K3.1 Pro scored 81.0% on the test, compared with 90.5% for an equivalent version without the risk-reduction measures.
• The test involved finding and exploiting software vulnerabilities in a controlled environment.
• Moonshot says the result demonstrates that the filtering reduced the model's offensive performance, although it still surpasses its predecessor.
• The results were not independently verified.

The challenges of open AI

Despite the measure, the Chinese company acknowledges that open models present specific control challenges. Unlike closed systems, such as those from OpenAI and Anthropic, since the files are widely distributed, the original developer has less capacity to restrict access, monitor use, or update safeguards than they would with a centrally hosted service.

Moonshot argues, on the other hand, that open models can contribute to cybersecurity by allowing more organizations to inspect and protect their own software, including older systems that receive little attention.

Concern about the use of AI in cyberattacks has recently taken on regulatory dimensions. Last week, China proposed export controls that would prohibit sending AI-based malware generation technologies abroad, as well as restrict artificial intelligence-assisted vulnerability discovery tools.

The measures cite the risk that such capabilities could be used against critical infrastructure, industrial and governmental defense systems, or to circumvent security protections.