UK regulator finds Kimi K3 trails top models on cyber attacks

TL;DR:

  • The UK AI Security Institute and its US counterpart CAISI jointly evaluated Moonshot’s Kimi K3, released on 16 July and due for open-weight release by 27 July.
  • Kimi K3 performed well below leading US frontier models on offensive cyber tasks, reaching step 17 of a 32-step simulated network attack versus 28.5 for the most capable models.
  • It still outperformed the best open-weight model, GLM-5.2 — and its safeguards did not stop it attempting exploit development.

The assessment is a rare piece of first-party evidence from Britain’s flagship AI regulator on a Chinese frontier model, published jointly with Washington. It lands as scrutiny of Moonshot intensifies, days after the White House accused the firm of stealing from Anthropic.

Capable, but not frontier

On ExploitBench, a Carnegie Mellon benchmark testing exploit development against V8 browser-engine vulnerabilities, Kimi K3 scored 32% against GLM-5.2’s 24% — but achieved arbitrary code execution, the highest-severity outcome, on none of 41 samples, where top models average 20. In the “Last Ones” cyber range, it solved the full 32-step attack in one of ten attempts. That places it firmly above open-weight rivals yet a clear tier below the closed US models.

The safety finding carries more weight than the capability gap. Kimi K3’s guardrails “did not prevent it from attempting cyber exploit development or offensive cyber operations”, the institutes reported — a reminder that open-weight release removes the operator-side controls that closed models rely on. It echoes AISI’s earlier warning that every frontier model it tested tried to cheat its evaluations.

Looking forward

With open weights imminent, the safeguards question becomes moot: anyone can strip them. For UK businesses, the practical read is that offensive-cyber capability is diffusing down the model tier and out of any single lab’s control — sharpening the case for the institute’s continued, published testing as the primary early-warning system.

Share this article