Moonshot AI Curbs Offensive Cyber Skills

Beijing-based Moonshot AI announced on Saturday that it has reduced the cyber-offensive capabilities of its newest AI model, Kimi K3.1 Pro. This move addresses widespread fears that increasingly autonomous AI could accelerate and simplify cyberattacks. The company intends to publish the model's weights, allowing external developers to run and modify it, by November 4.

Moonshot stated that it removed cyber-related data from the model's training corpus and reinforcement learning environments, while preserving its general coding skills. This decision aims to prevent the release of open-weight models that would advance the frontier of cyber capabilities. The startup believes open-weight models can improve overall security by enabling more organizations to inspect and secure their software.

This development occurs as governments and security experts express growing concern over AI's potential to make sophisticated hacking more accessible. AI models capable of writing code and performing multi-step tasks could aid in identifying software weaknesses and automating cyberattack components. In response, China recently proposed export controls that would ban AI-driven malware technology and restrict AI-assisted vulnerability discovery tools.

K3.1 Pro demonstrated a score of 81.0% on an internal cyber benchmark, a decrease from a similar version without risk-reduction measures which scored 90.5%. This test involved exploiting software vulnerabilities in a controlled setting. While the results have not been independently verified, Moonshot indicated that filtering cyber-specific data effectively reduced the model's offensive cyber performance. The model also features a 1-million-token context window for processing large text volumes and a faster high-speed version is available.