Moonshot Limits Hacking Skills in Kimi K3.1 Pro AI Model

Chinese AI startup Moonshot AI is deliberately limiting the hacking capabilities of its newest model, putting a spotlight on a dilemma facing the whole industry: how to make AI more powerful without making cyberattacks easier.

The company introduced Kimi K3.1 Pro on October 24, describing it as a highly capable open-weight model with around 3 trillion parameters. Open-weight models let developers download the model files, run them on their own systems and modify how they behave. Moonshot plans to release the weights on November 4.

Cyber data removed from training

According to Moonshot, it removed cyber-related data from the model's training dataset and reinforcement-learning environments to reduce its ability to carry out offensive cyber operations, while keeping general coding skills intact.

The company's internal tests suggest the approach works. Kimi K3.1 Pro scored 81% on Moonshot's cybersecurity benchmark, which involves finding and exploiting software vulnerabilities in a controlled environment. A similar version without the risk-reduction measures scored 90.5%, a gap of 9.5 percentage points. These results have not been independently verified, and the model reportedly still outperformed its predecessor.

The model also offers a one-million-token context window, and Moonshot has added a Highspeed version that it says runs two to three times faster than the standard model.

Why it matters

The same capabilities can serve defence and offence. Security teams can use AI to find weaknesses before criminals do, while attackers can use models that write code, operate tools and complete multi-step tasks to automate parts of an attack. Moonshot argues open-weight AI can help organisations inspect and secure neglected or older software.

Once model files are widely distributed, though, the original developer has less control over who uses them and whether safeguards hold. The issue is also reaching policymakers: China has proposed export controls on AI-driven malware-generation technology and certain AI-assisted vulnerability-discovery tools, and those proposals remain open to public comment.

The next test arrives on November 4, when developers and security researchers gain direct access to the weights. Then the industry will learn whether safeguards built into training can survive independent use.