NewsBite analysis · AI, Security
The Decoder
Kimi K3 trails frontier US models by a wide margin on cyber exploits, and distillation may explain why
The Kimi K3 model, developed by Moonshot AI, fell short in cybersecurity evaluations conducted by UK and U.S. AI institutes, notably trailing leading U.S. models in offensive cyber task performance. While it surpassed China's GLM-5.2, major gaps in exploit capabilities were evident, raising alarm about operational risks in software exploitation. Kimi K3's training may have limited its capacity to handle advanced cyber tasks as it largely relies on general dataset learnings, implying underlying vulnerabilities.
MonitorConfidence: HighShort term horizonCybersecurity risk: High
Published 24 Jul 2026, 10:10 UTC3 min read
Visit original article ↗Lead image — The Decoder
Summary
TL;DR
- Kimi K3 scored 32.2% in cyber exploit tasks, lagging behind U.S. models at 76.2%.
- The model failed to achieve Arbitrary Code Execution (ACE) in any tests.
- Kimi K3 reached only step 17 out of 32 in a simulated corporate network attack.
- Concerns over the potential misuse of Kimi K3's capabilities were raised.
- U.S. models showed a growing performance gap over Chinese counterparts.
Interesting facts
- Kimi K3's performance may have been influenced by distillation from higher-performing models.
- The study used a benchmark developed by Carnegie Mellon University focused on exploit development.