Beijing-based AI startup Moonshot AI has reduced the cyberattack capabilities of its latest system, Kimi K3.1 Pro. Revealed on October 24, the update precedes the planned November 4, 2026 release of the model’s weights, arriving amid escalating concerns that artificial intelligence could lower the cost and complexity of launching cyberattacks.
What Moonshot Changed
Moonshot eliminated cyber-related information from the model’s training datasets and reinforcement-learning environments. According to the company, the system’s overall programming abilities remain intact. This adjustment aims to prevent the distribution of an open-weight model that substantially enhances the hacking proficiency of malicious actors.
Hailed by the firm as its most powerful system to date, Kimi K3.1 Pro features three trillion parameters. The model’s core files, or weights, will be made available for developers to download, enabling them to operate or modify the software on their own infrastructure. Furthermore, the system processes up to one million tokens concurrently, alongside a high-speed variant, K3.1 Pro Highspeed, that operates two to three times faster.
During internal evaluations, the model achieved an 81% success rate in identifying and exploiting software vulnerabilities within a controlled environment. By comparison, a version lacking these safety measures scored 90.5%. Although independent experts have not yet verified these figures, Moonshot maintains that the new iteration outperforms its predecessor.
Also read: Chinese AI Model Kimi K3 Bypasses UK Cybersecurity Sandbox During Testing
Why Open Models Worry Experts
Security researchers, technology corporations, and government bodies worry that advanced AI tools might assist threat actors in discovering software bugs and executing sophisticated attacks with minimal human oversight. Systems capable of generating code, utilizing software utilities, and executing multi-step tasks can reduce the technical expertise required to mount such operations.
Last week, China introduced proposed export controls focusing on artificial intelligence tools designed to build or alter malware. These proposed rules also seek to restrict AI-assisted vulnerability discovery to safeguard critical infrastructure, government networks, and defense organizations.
Moonshot acknowledges that open models assist organizations in identifying security gaps and defending legacy systems. However, open weights present greater oversight challenges than AI hosted directly on a vendor’s servers. Once released, these weights can be freely duplicated, modified, and redistributed, leaving the original creator unable to enforce usage restrictions.
What Comes Next
Moonshot’s decision highlights the ongoing difficulty of reconciling public access with safety precautions. While internal testing indicates the modifications were successful, independent verification is still required. True validation may arrive only after developers receive the model weights following the November 4 release and conduct their own evaluations.
