TLDR Tech / InfoSec
UK AISI/CAISI Preliminary Assessment of Kimi K3's Cyber Capabilities
UK AISI and CAISI jointly evaluated Moonshot AI's Kimi K3, released July 16 and slated for open-weight release by July 27, and placed it well behind frontier US models while ahead of GLM-5.2, scoring 32% on Carnegie Mellon's ExploitBench against 76.2% for top US models and 24.4% for GLM-5.2, and reaching step 17 of the 32-step "The Last Ones" cyber range against 28.5 steps for the leading US models. The gap narrows unevenly because Kimi K3 matched the frontier on coverage and came close on bug reproduction but achieved arbitrary code execution on zero of 41 samples where top models managed 20, and yet it completed the full cyber range once in ten attempts, which the evaluators read as autonomous capability against small, weakly defended enterprise networks given initial access. Two framing points deserve weight, the first being that Kimi K3's safeguards did not stop it from attempting exploit development or offensive operations at all, and the second being that US closed-weight comparators ran with system-level safeguards disabled to measure maximum capability, so the published gap measures raw model ability rather than what a user encounters, a distinction that matters as open-weight releases from PRC labs keep closing the trendline.
Read the full story