NEWS / OCT.2026
Moonshot says it has limited Kimi K3.1 Pro’s cyberattack capabilities
Moonshot says it has reduced Kimi K3.1 Pro’s capabilities that could be used for cyberattacks by changing its training. In an internal test, the company reports that 81% of exercises were completed successfully within a maximum of three attempts after filtering, compared with 90.5% without those measures.

Kimi K3.1 Pro: Moonshot says it has limited its cyber capabilities
On October 23, 2026, Moonshot introduced Kimi K3.1 Pro, saying it had limited its new model’s cyberattack capabilities. The company says it modified part of the model’s training for this purpose.
Moonshot says it removed cybersecurity data from intermediate training and excluded cyber tasks from reinforcement learning environments. During this phase, the model practices using rewards. General programming tasks were retained.
Searching for vulnerabilities can help security teams audit their software and prepare fixes. Attackers can also use these same capabilities.
A vulnerability test in an isolated environment
To assess the effect of filtering, Moonshot uses xdev-bench-hard, a new internal test dedicated to discovering and exploiting vulnerabilities. The exercises provide, among other things, vulnerable code, a compiled program and a hint. The model must exploit the vulnerability in an isolated environment.
Moonshot reports that Kimi K3.1 Pro successfully completed 81% of exercises after filtering, compared with 90.5% for a comparable version without these measures and 33.3% for Kimi K3. The pass@3 score counts a task as successful if it is completed successfully at least once in three attempts.
The difference between the two versions of Kimi K3.1 Pro amounts to 9.5 percentage points. Some exercises are still completed successfully after filtering, under the test conditions. The test takes place in an isolated environment and does not measure the success of real-world cyberattacks.
An independent defensive test by Artificial Analysis gives Kimi K3.1 Pro a score of 64 on its Cyber Index. It assesses the discovery, validation and remediation of vulnerabilities, without measuring the creation of exploit programs. These tasks differ from those in the internal test and cannot verify the reduction reported by Moonshot.
Moonshot says Kimi K3.1 Pro is available in its Kimi products and through its application programming interface, or API. The company promises to release the model’s full weights by November 4, 2026. These parameters would allow it to be run locally.