Beijing startup Moonshot removes hacking data from new AI model

Moonshot has taken deliberate steps to constrain the hacking capabilities of its newly developed Kimi artificial intelligence model, signalling growing industry concern over potential security risks posed by advanced AI systems. The startup removed cyber-related data from both the model's training materials and testing environments, according to statements from the company.

The move reflects broader efforts within China's AI sector to implement safety measures and responsible development practices. As large language models become increasingly sophisticated, companies face mounting pressure to ensure their systems cannot be weaponised for malicious purposes such as orchestrating cyber attacks or exploiting software vulnerabilities.

Moonshot's approach involves filtering out training data related to hacking techniques and preventing the model from being tested in scenarios that would reinforce cyber-attack capabilities. This represents a proactive step beyond standard safety protocols typically employed by AI developers.

The decision comes amid heightened scrutiny of AI development practices globally, with regulators and security experts warning that unrestricted AI systems could pose significant cybersecurity threats. Other major AI developers have similarly implemented guardrails to prevent their models from being used for harmful purposes.