The UK Artificial Intelligence Security Institute (UK AISI) and U.S. Center for AI Standards and Innovation (CAISI) jointly evaluated Moonshot AI's Kimi K3, released July 16, 2026 with open-weight release slated for July 27, 2026. The preliminary assessment focused on cyber capabilities using Exploit Bench and The Last Ones (TLO) cyber range.

On Exploit Bench's 41 exploit development tasks, Kimi K3 achieved a 32% success rate, outperforming GLM-5.2's 24%. However, Kimi K3 failed to achieve arbitrary code execution (ACE) on any samples (0/41), while top models averaged 20/41 ACE successes.

In the 32-step TLO cyber range simulating corporate network attacks, Kimi K3 reached step 17 on average versus 28.5 for leading U.S. models, but surpassed GLM-5.2's step 11. Kimi K3 completed the full range in 1 of 10 attempts within the 100M-token limit, indicating capability against weakly defended systems.

Evaluators note TLO lacks active defenders and penalties for triggering alerts. Four closed-weight models have previously solved TLO, with top models solving 6/10 or 7/10 attempts reliably.