The Beijing-based company announced the changes on Saturday, October 24, 2026, ahead of plans to release the model’s underlying weights on November 4.
These files allow developers to run, modify and adapt an AI model on their own computer systems.
According to Reuters, Moonshot removed cybersecurity-related data from the model’s training material and reinforcement-learning environments while retaining its general coding capabilities.
The company said the decision was intended to prevent the public release of a model that could significantly expand the ability to conduct sophisticated cyberattacks.
“We aim not to release open weights that would extend the absolute frontier of cyber capability,” Moonshot said.
The move comes as governments, technology companies and cybersecurity researchers increasingly warn that advanced AI systems could help malicious actors identify software weaknesses and automate parts of cyberattacks.
AI models capable of writing code, using software tools and completing complex tasks could make it easier to discover vulnerabilities in computer systems, including those supporting critical infrastructure and government services.
Moonshot said its risk-reduction measures had lowered the model’s performance on an internal cybersecurity benchmark.
The company reported that Kimi K3.1 Pro scored 81.0% on a controlled test involving the identification and exploitation of software vulnerabilities, compared with 90.5% for a similar version without the additional restrictions.
The results have not been independently verified.
Despite the reduction, Moonshot said the new model still outperformed its predecessor.
The company maintains that openly available AI models can also improve cybersecurity by helping organisations identify weaknesses and protect their software, particularly older systems that may not receive regular security updates.
However, releasing model weights publicly presents additional challenges because developers have less control over how the technology is used once the files are widely distributed.
The announcement follows China’s proposed export controls on certain AI-driven malware-generation technologies and tools that use AI to identify software vulnerabilities.
The proposed measures reflect growing concerns about the potential misuse of increasingly capable AI systems.
Kimi K3.1 Pro features a context window of one million tokens, allowing it to process large amounts of information in a single interaction.
Moonshot also introduced a faster version, K3.1 Pro Highspeed, which the company says operates two to three times faster than the standard model.
Leave a Reply