UK and US Jointly Assess Kimi K3's Cyber Capabilities
Preliminary evaluation reveals AI model's cybersecurity performance in benchmark tests.
1Pioneering Evaluation Report
UK AISI and US CAISI conducted tests on Moonshot AI's Kimi K3 model's cyber capabilities. Released on July 16, 2026, the model will be available as open-weight on July 27. The assessment focused on measuring the model's exploit development capacity through specialized benchmarks.
2ExploitBench Performance
Kimi K3 achieved 32% success rate in Carnegie Mellon's ExploitBench, outperforming June 2026's most capable open-weight model GLM-5.2 (24%). However, the model failed in the most critical cybersecurity scenario - arbitrary code execution, which represents the highest severity exploit outcome.
3Cyber Range Test Results
In the 32-step 'The Last Ones' cyber range simulation, Kimi K3 reached step 17 on average. While top US models completed 28.5 steps, Kimi succeeded in fully completing the range in 1 out of 10 attempts, demonstrating potential to attack weakly defended systems when provided initial access.
4Comparative Analysis
The report compares Kimi K3's performance with global models. US closed-weight models showed superior capabilities when safety restrictions were removed. However, Kimi K3 emerged as a leading performer among Chinese open-weight models in cybersecurity benchmarks.
5Methodology and Limitations
Assessments used Item Response Theory-inspired approach. Kimi K3's evaluation scope was narrower than other models. Researchers noted cyber range tests don't fully replicate real-world conditions as they lack active defense mechanisms and intentional alert systems.
6Industry Significance
This evaluation establishes critical reference points for measuring AI models' cybersecurity capabilities. It provides valuable insights for developers and security professionals in understanding the offensive potential of open-weight AI models in controlled environments.
Original article: Hacker News (news.ycombinator.com) — Kimi K3'nin Siber Yetenekleri Değerlendirildi