英美AI安全机构联合评估Kimi K3网络能力

英国人工智能安全研究所(UK AISI)与美国AI标准创新中心(CAISI)近日联合发布了对月之暗面公司Kimi K3模型网络能力的初步评估报告。1

评估结果

在漏洞利用基准测试(ExploitBench)中，Kimi K3获得32%的得分，超过了同类开源模型GLM-5.2的24%。1该测试涵盖了41个2023年后发现的V8引擎漏洞。1

但在高级网络攻击模拟测试"The Last Ones"(TLO)中，Kimi K3的表现明显低于顶尖美国模型。1根据测试数据，Kimi K3在该攻击链中平均完成至第17步(总共32步)，而美国顶尖模型平均完成28.5步。1在10次尝试中，Kimi K3仅成功完成1次。1

在任意代码执行(ACE)能力方面，Kimi K3在41个样本测试中达成0次，而最顶尖模型平均达成20次。1评估结果表明Kimi K3未能在任何任务中实现任意代码执行。1

产品发布计划

Kimi K3已于2026年7月16日发布，计划于2026年7月27日发布开源版本。1

---

The UK Artificial Intelligence Safety Institute (UK AISI) and the US Center for AI Safety and Innovation (CAISI) have jointly evaluated the cyber capabilities of Kimi K3, a model developed by Moonshot AI. 1

The assessment measured performance across multiple cybersecurity benchmarks. On the ExploitBench vulnerability exploitation benchmark, Kimi K3 achieved a score of 32%, outperforming the open-source GLM-5.2 model, which scored 24%. 1 However, in advanced attack simulation testing known as "The Last Ones" (TLO), Kimi K3 performed below leading American models. 1

Notably, Kimi K3 failed to achieve arbitrary code execution (ACE) in any test, registering zero successful instances across 41 samples, while leading American models averaged 20 successful ACE instances. 1 In the TLO attack chain assessment, Kimi K3 completed an average of step 17 out of 32 total steps, compared to an average of step 28.5 for top American models. 1 The model achieved only one successful completion in 10 attempted runs of the TLO scenario. 1

The ExploitBench testing covered 41 V8 engine vulnerabilities discovered after 2023. 1 Kimi K3 was released on July 16, 2026, with an open-source version scheduled for release on July 27, 2026. 1

---