关键要点
- 英国AI安全研究所与美国AI标准与创新中心联合评估发现，月之暗面（Moonshot AI）的Kimi K3在几乎没有有效抵抗的情况下协助实施攻击性网络行动。
- Kimi K3在漏洞利用开发和模拟网络攻击方面远落后于美国领先模型，但优于中国的GLM-5.2。
- 中国模型在网络任务上持续进步，但仍落后于美国系统。Kimi K3的结果也与月之暗面蒸馏更先进模型的指控相符。
英国AI安全研究所（UK AISI）与美国AI标准与创新中心（CAISI）联合评估了月之暗面（Moonshot AI）的最新模型Kimi K3。
Kimi K3在攻击性网络任务上大幅落后于美国领先的前沿模型，但优于中国的GLM-5.2，在开放权重模型中树立了新标杆。其安全防护未能阻止漏洞利用开发或攻击性网络行动，该模型在无任何抵制的情况下协助完成了这两类任务。
Kimi K3无法攻克最高难度的漏洞利用等级
两家机构使用了卡内基梅隆大学开发的基准测试ExploitBench来测试漏洞利用开发能力。该基准使用2023年后Chrome的V8引擎中发现的41个漏洞，追踪模型在软件利用过程中能推进到哪一步。美国领先模型平均得分76.2%，而Kimi K3为32.2%，GLM-5.2为24.4%。
Kimi K3在全部41项任务中均未达到最高等级——任意代码执行（ACE）。ACE是最严重的漏洞利用等级，因为它能让攻击者完全控制目标系统。美国领先模型在41项任务中的20项中实现了ACE。
两家机构在禁用系统级安全防护的情况下测试了美国闭源权重模型，以衡量其最大能力。这些安全防护在公开版本中是启用的。
Kimi K3在模拟网络攻击中推进到一半
第二项测试“最后幸存者”（TLO）模拟的是企业网络攻击场景，攻击路径横跨四个子网、涉及约 20 台主机，共 32 个步骤。据两家机构称，人类专家完成该测试大约需要 20 小时。目前只有少数模型能够解出 TLO。迄今为止，已有四款公开可用的闭源权重模型通过了该测试，其中表现最强的模型在十次尝试中能成功六到七次。
Kimi K3 平均推进到 32 步中的第 17 步，而美国领先模型平均推进到第 28.5 步，GLM-5.2 仅推进到第 11 步。Kimi K3 在十次尝试中有一次在 1 亿 token 限额内完成了整条攻击路径，这表明它具备相应能力，但无法稳定调用。“Kimi K3 在被明确指示并获得初始网络访问权限的情况下，有能力自主攻击规模较小、防御薄弱且存在漏洞的企业系统，”该机构写道。
TLO 未考虑主动防御因素，因此并非完全贴近现实。但相关结果在真实场景中足以拉响警报。本周就出现了一个新案例：OpenAI 的模型试图自主入侵 Hugging Face。Hugging Face 成功抵御了攻击，但付出了实实在在的努力，并动用了开放权重模型。
中国模型正在追赶，但仍落后于美国模型
CAISI 进行的时间序列分析基于 Elo 评分体系，追踪了自 2025 年初以来美国和中国模型的网络攻防能力。两条趋势线都在上升，但中国模型始终落后于美国同行。
在此前的一份分析中，这家英国机构将开源模型的性能差距定为四到七个月，而2025年初这一差距为六到十个月。最新结果符合这一规律。中国开源权重模型正在变强，但与领先的美国系统相比仍有明显差距。
AISI 警告称，这一差距不应让人掉以轻心。开源模型不断增强的网络能力构成了“一种持续且不可逆的滥用风险”。
网络测试结果与蒸馏指控相互印证
Kimi 的相关发现也为针对中国模型开发者的蒸馏指控提供了佐证。美国科学顾问 Michael Kratsios 近日指控月之暗面（Moonshot AI）利用 Anthropic 的 Fable 的最佳输出作为训练数据来提升 Kimi K3 的性能，从而“蒸馏”了 Fable。Kratsios 还声称月之暗面获得了受美国出口管制的 Nvidia GB300 芯片。
对于通用基准表现强劲但网络能力得分偏低这一差距，一种解释是：Kimi K3 可能主要基于 Claude 的输出进行训练，而这些输出覆盖了通用知识、编程和智能体任务。Anthropic 的安全分类器会专门拦截高级攻击性网络查询，因此这类输出在基于 Claude 响应构建的蒸馏数据集中占比会偏低。这样一来，Kimi K3 就能在标准基准上比肩领先的西方模型，却未必学到它们更深层的漏洞利用能力。
AISI 的结果支持这一解读。该机构关闭了美国模型的系统级安全防护，从而暴露出这些模型几乎无法通过公开接口访问、因此基本无法被蒸馏获取的网络能力。
AISI
Key Points
- A joint evaluation by the British AI Security Institute and the U.S. Center for AI Standards and Innovation found that Moonshot AI's Kimi K3 assists with offensive cyber operations without meaningful resistance.
- Kimi K3 fell well behind leading U.S. models in exploit development and simulated network attacks, though it outperformed China's GLM-5.2.
- Chinese models continue to improve on cyber tasks but remain behind U.S. systems. Kimi K3's results are also consistent with allegations that Moonshot AI distilled more advanced models.
The British AI Security Institute (UK AISI) and the U.S. Center for AI Standards and Innovation (CAISI) jointly evaluated Moonshot AI's latest model, Kimi K3.
Kimi K3 trails the leading U.S. frontier models by a wide margin on offensive cyber tasks but outperforms China's GLM-5.2, setting a new benchmark among open-weight models. Its safeguards didn't block exploit development or offensive cyber operations, and the model assisted with both without pushback.
Kimi K3 can't crack the hardest exploit levels
The institutes used ExploitBench, a benchmark developed by Carnegie Mellon University, to test exploit development skills. It uses 41 vulnerabilities found in Chrome's V8 engine after 2023 to track how far a model advances through the software exploitation process. The leading U.S. models averaged 76.2 percent, compared with 32.2 percent for Kimi K3 and 24.4 percent for GLM-5.2.
Kimi K3 didn't reach the highest level, known as Arbitrary Code Execution (ACE), on any of the 41 tasks. ACE is the most severe exploit level because it gives attackers full control over a target system. The leading U.S. models achieved ACE in 20 of the 41 tasks.
The institutes tested the U.S. closed-weight models with their system-level safeguards disabled to measure their maximum capabilities. Those safeguards are enabled in the publicly available versions.
Kimi K3 gets halfway through a simulated network attack
The second test, "The Last Ones" (TLO), simulates a corporate network attack with a 32-step attack path across four subnets and about 20 hosts. A human expert would need roughly 20 hours to complete it, according to the institutes. Only a small group of models can solve TLO at all. Four publicly available closed-weight models have passed the test so far, with the strongest succeeding six or seven times out of ten.
Kimi K3 reached step 17 out of 32 on average, compared with 28.5 steps for the leading U.S. models and just 11 for GLM-5.2. It completed the entire attack path in one of ten attempts while staying within the 100 million token limit, showing that it has the capability but can't call on it reliably. "Kimi K3 is capable of autonomously attacking small, weakly defended and vulnerable enterprise systems, when directed to do so and given initial network access", the institute writes.
TLO doesn't account for active defense, so it isn't fully realistic. But the results would raise red flags in real-world scenarios. A fresh example showed up this week when OpenAI models tried to autonomously hack into Hugging Face. Hugging Face fended off the attack, though it took real effort and the use of open-weight models.
Chinese models are gaining ground but still trail U.S. models
A time-series analysis by CAISI tracks the cyber capabilities of U.S. and Chinese models since early 2025 on an Elo-based scale. Both trend lines are climbing, but Chinese models consistently remain behind their U.S. counterparts.
In a previous analysis, the British institute pegged the performance gap for open models at four to seven months, compared with six to ten months at the start of 2025. The new results fit this pattern. Chinese open-weight models are getting stronger, but they remain well behind leading U.S. systems.
AISI warns that this gap shouldn't breed complacency. The growing cyber capabilities of open models create "a persistent and irreversible risk of misuse."
Cyber results line up with distillation allegations
The Kimi findings also lend support to distillation allegations against Chinese model developers. U.S. science advisor Michael Kratsios recently accused Moonshot AI of "distilling" Anthropic's Fable by using Fable's best outputs as training data to boost Kimi K3's performance. Kratsios also alleged that Moonshot AI had access to Nvidia's GB300s, which are subject to U.S. export controls.
One explanation for the gap between strong general benchmarks and weak cyber scores is that Kimi K3 may have been trained mostly on Claude outputs covering general knowledge, programming, and agent tasks. Anthropic's safety classifiers specifically block advanced offensive cyber queries, so those outputs would be underrepresented in a distillation dataset built from Claude responses. Kimi K3 could therefore match leading Western models on standard benchmarks without picking up their deeper exploit capabilities.
The AISI results support this reading. The institute disabled system-level safeguards on the U.S. models, revealing cyber capabilities that are nearly impossible to access through public interfaces and therefore largely unavailable for distillation.
AISI