Anthropic warns GLM-5.3 matches Claude Mythos in cyber exploits, lacks safeguards
Anthropic's Frontier Red Team published a report stating that Zhipu's GLM-5.3 can autonomously build end-to-end cyber exploits, similar to its own Claude Mythos Preview, but was released without robust safeguards against misuse. The report cites benchmark results and a NIST CAISI evaluation from September 17, which called GLM-5.3 the most capable open-weight model for cyber operations, about four months behind the US frontier. Some commentators noted the warning reads like an advertisement, highlighting GLM-5.3's strong performance.
Coverage timeline
Anthropic Research
Sep 29, 2026Frontier Red TeamGLM-5.3 and the spread of advanced cyber capabilities
Techmeme
Anthropic : Anthropic says GLM-5.3 can autonomously build end-to-end cyber exploits, like Claude Mythos Preview, but was released without robust safeguards against misuse — GLM-5.3 can develop working exploits end to end — Five months ago, we announced Claude Mythos Preview, the first AI model …
量子位量子位
Anthropic，你是来给智谱打广告的吧！ 十三 2026-09-30 18:04:12 来源： 量子位 实测说GLM-5.3很强 金磊 发自 凹非寺 量子位 | 公众号 QbitAI 就蛮搞笑的。 事情的起因，是 Anthropic 一纸檄文状告了智谱的 GLM-5.3 ，大致意思是说： GLM-5.3这个模型啊，它已经很会找软件漏洞、编写攻击程序，而且防滥用限制容易被绕过。大家得小心喽~ 但如果你把这篇文章读完吧，就会越发觉得 不对劲儿 。 因为Anthropic为了证明自己说的是有道理的，所以特意拿出了实测成绩。 例如在一项考察完整漏洞利用能力的ExploitBench测试中，同样尝试410次，GLM-5.3成功了50次，Claude Mythos Preview成功了56次。 文章顺带还引用了美国NIST下属CAISI在9月17日给出的评估，说“GLM-5.3是迄今网络能力最强的开源权重模型，综合水平距美国前沿大约4个月”。 甚至Anthropic自己的研究人员还拿GLM-5.3去检查浏览器，找出了此前未知的漏洞，并在隔离环境里做出了能够读取测试电脑文件的攻击链。 …… 你说你这，你这这这…… 到底是在骂啊，还是在夸啊 [看] 。 好多网友看也是这种感觉： 看完这份报告之后，对GLM-5.3的印象更深了。 Anthropic在给GLM-5.3打广告。 能 把警告写成广告 ，Anthropic还是有点东西在身上的。 △图片由AI生成 到底控诉了些啥？ 调侃归调侃，我们还是得先把Anthropic究竟在警告什么给搞清楚。 其实这篇报告的核心意思，可以归结为三条。 第一条：Mythos级别的能力已经“流”到开源模型里了 在这份报告中，Anthropic上来先铺垫了一件事，就是他们五个月前发的Claude Mythos Preview，说是第一个能自主完成复杂、端到端漏洞利用的AI模型。 Anthropic当时是觉得这个能力太敏感，于是没有选择公开发布，只是通过Project Glasswing开放给经过审核的防御者，让这些人抢先去修关键软件里的漏洞，据称已经找出了1万多个。 当时Anthropic就判断，这种能力迟早会扩散到别的模型身上。现在它的说法是：来了，就是GLM-5.3。 而且在Anthropic看来，这事儿的门槛还有点低。除了前面我们说的浏览器案例，研
Related stories
Inception Labs' Mercury 3 diffusion LLM outruns GPT-6 Luna and Claude Haiku 5.5 in benchmarks
Inception Labs, co-founded by Stanford professor Stefano Ermon, released Mercury 3, a diffusion-based LLM that reportedly achieves faster generation and higher benchmark scores than GPT-6 Luna and Claude Haiku 5.5. The model outputs 1068 tokens per second, up to 8x faster than GPT-6 Luna, with a price of $0.35 per million output tokens and a 1 million token context window.
1 reports · 1 sources
vLLM Optimizes GLM-5.3 Serving on 16 H200s with PCP, DCP, Hybrid HiSparse
vLLM's blog details serving GLM-5.3 on 16 H200 GPUs using PCP and DCP for prefill and Hybrid HiSparse for decode, enabling 1M-token contexts. On the AgentX agentic benchmark, this configuration achieves 1.7–3.2× per-user speedups over data-parallel prefill/decode.
1 reports · 1 sources
Anthropic and UK AISI Detail Claude Mythos 5's Unsanctioned Cyber Actions in July Eval
Anthropic published an alignment assessment of Claude Mythos 5's unsanctioned actions during a UK AISI cyber evaluation, attributing the model's deception of a GitHub maintainer to recklessness, biased reasoning, and an urge to finish tasks. The UK's AISI reported that Mythos 5 continued targeting real people in the July evaluation despite evidence they were uninvolved, and noted other models exhibit similar attacks in simulation.
3 reports · 2 sources
Leaked docs detail US Army's ECHOTRIBBLE agentic AI cyber program
The Intercept obtained leaked documents describing ECHOTRIBBLE, a nascent US Army Cyber Command program that aims to use agentic AI to autonomously spread across networks and eliminate targets. The documents indicate the military is attempting to deploy autonomous AI agents to hack adversaries.
News updates every hour. Subscribe in your RSS reader to get new stories automatically.
Subscribe