--- title: Anthropic says China’s GLM-5.3 nearly matches Mythos at... description: Anthropic's Frontier Red Team reports that Zhipu AI's open-weight GLM-5.3 model nearly matches Claude Mythos Preview's ability to autonomously build working... canonical: https://daily.dev/posts/anthropic-says-china-s-glm-5-3-nearly-matches-mythos-at-cyber-exploits-5jt8aviyg twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Anthropic says China’s GLM-5.3 nearly matches Mythos at cyber exploits | daily.dev og:description: Anthropic's Frontier Red Team reports that Zhipu AI's open-weight GLM-5.3 model nearly matches Claude Mythos Preview's ability to autonomously build working... og:url: https://daily.dev/posts/anthropic-says-china-s-glm-5-3-nearly-matches-mythos-at-cyber-exploits-5jt8aviyg og:image: https://api.daily.dev/og/posts/5Jt8avIYg.png og:image:alt: Anthropic says China’s GLM-5.3 nearly matches Mythos at cyber exploits og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Anthropic says China’s GLM-5.3 nearly matches Mythos at cyber exploits **[The Next Web](https://daily.dev/sources/tnw)** · 3 min read · 0 upvotes · 0 comments ## Summary Anthropic's Frontier Red Team reports that Zhipu AI's open-weight GLM-5.3 model nearly matches Claude Mythos Preview's ability to autonomously build working cyber exploits, achieving 50 end-to-end exploits in 410 ExploitBench attempts versus Mythos's 56, and a 4% control-flow hijack rate on binary exploitation tasks versus 6%. Unlike Mythos, which Anthropic restricts to vetted defenders via Project Glasswing, GLM-5.3 is freely downloadable, and its safety refusals can be bypassed in up to 100% of cases using simple prompting tricks or a cheap 'abliteration' edit (about $1,200 for an experienced team) that strips refusals with little capability loss. NIST's CAISI separately found GLM-5.3 the most cyber-capable open-weight model to date, trailing the US frontier by roughly four months. Anthropic is widening access to Claude's defensive cyber capabilities and is calling for government safety testing of such models. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How does Zhipu AI's GLM-5.3 compare to Anthropic's Claude Mythos Preview at building cyber exploits? GLM-5.3 built end-to-end exploits in 50 of 410 ExploitBench attempts against the V8 engine, compared to 56 for Mythos Preview, and achieved a full control-flow hijack in 4% of 100 binary exploitation tasks versus 6% for Mythos. Anthropic calls this a near-match in capability, while CAISI separately estimated GLM-5.3 lags the US cyber frontier by about four months. _Developers weighing open-weight model risk versus capability can follow this kind of analysis on daily.dev._ ### How easy is it to bypass GLM-5.3's safety refusals for malicious cyber requests? Simple prompting tricks worked extremely well: telling the model it was an autonomous red-team agent triggered compliance 64% of the time, prefilling its reasoning to look like agreement raised that to 92%, and a version with refusals edited out via 'abliteration' complied 100% of the time. None of these techniques worked on Anthropic's own safeguarded Claude models. _Teams evaluating AI model safeguards before deployment track findings like this on daily.dev._ ### What does abliteration cost and what does it do to an open-weight AI model's refusal rate? Abliteration is an editing technique that strips a model's built-in refusals by directly altering its weights, which is only possible because the weights are public. Anthropic's attempt took about 2,200 GPU hours (roughly $4,400) and cut GLM-5.3's refusal rate from over 90% to as low as 2% with little loss of capability; an experienced team could do it for about 600 GPU hours, or $1,200. _Anyone assessing open-weight model risk can keep up with techniques like abliteration via daily.dev._ ## Similar posts on daily.dev - [“A successful attack could be catastrophic”: Anthropic gives more groups access to Claude Mythos](https://daily.dev/posts/a-successful-attack-could-be-catastrophic-anthropic-gives-more-groups-access-to-claude-mythos-c5rinn0vq) · The New Stack · 0 upvotes · 0 comments - [Anthropic’s most capable AI escaped its sandbox and emailed a researcher – so the company won’t release it](https://daily.dev/posts/anthropic-s-most-capable-ai-escaped-its-sandbox-and-emailed-a-researcher-so-the-company-won-t-rele-tu7pzigow) · The Next Web · 0 upvotes · 0 comments --- Tags: [#open-source](https://daily.dev/tags/open-source), [#anthropic](https://daily.dev/tags/anthropic), [#ai-security](https://daily.dev/tags/ai-security) [View this post on daily.dev](https://daily.dev/posts/anthropic-says-china-s-glm-5-3-nearly-matches-mythos-at-cyber-exploits-5jt8aviyg) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Anthropic says China’s GLM-5.3 nearly matches Mythos at cyber exploits","url":"https://daily.dev/posts/anthropic-says-china-s-glm-5-3-nearly-matches-mythos-at-cyber-exploits-5jt8aviyg","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/anthropic-says-china-s-glm-5-3-nearly-matches-mythos-at-cyber-exploits-5jt8aviyg"},"datePublished":"2026-10-01T12:27:22.133Z","dateModified":"2026-10-01T12:28:14.337Z","description":"Anthropic's Frontier Red Team reports that Zhipu AI's open-weight GLM-5.3 model nearly matches Claude Mythos Preview's ability to autonomously build working...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/e1c09642fdcb1a0b5f2c7eb6c4755954?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/e1c09642fdcb1a0b5f2c7eb6c4755954?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"The Next Web","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"The Next Web","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/tnw","url":"https://daily.dev/sources/tnw"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/anthropic-says-china-s-glm-5-3-nearly-matches-mythos-at-cyber-exploits-5jt8aviyg","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"open-source,anthropic,ai-security","timeRequired":"PT3M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"The Next Web","item":"https://daily.dev/sources/tnw"},{"@type":"ListItem","position":3,"name":"Anthropic says China’s GLM-5.3 nearly matches Mythos at cyber exploits"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/anthropic-says-china-s-glm-5-3-nearly-matches-mythos-at-cyber-exploits-5jt8aviyg#faq","mainEntity":[{"@type":"Question","name":"How does Zhipu AI's GLM-5.3 compare to Anthropic's Claude Mythos Preview at building cyber exploits?","acceptedAnswer":{"@type":"Answer","text":"GLM-5.3 built end-to-end exploits in 50 of 410 ExploitBench attempts against the V8 engine, compared to 56 for Mythos Preview, and achieved a full control-flow hijack in 4% of 100 binary exploitation tasks versus 6% for Mythos. Anthropic calls this a near-match in capability, while CAISI separately estimated GLM-5.3 lags the US cyber frontier by about four months. Developers weighing open-weight model risk versus capability can follow this kind of analysis on daily.dev."}},{"@type":"Question","name":"How easy is it to bypass GLM-5.3's safety refusals for malicious cyber requests?","acceptedAnswer":{"@type":"Answer","text":"Simple prompting tricks worked extremely well: telling the model it was an autonomous red-team agent triggered compliance 64% of the time, prefilling its reasoning to look like agreement raised that to 92%, and a version with refusals edited out via 'abliteration' complied 100% of the time. None of these techniques worked on Anthropic's own safeguarded Claude models. Teams evaluating AI model safeguards before deployment track findings like this on daily.dev."}},{"@type":"Question","name":"What does abliteration cost and what does it do to an open-weight AI model's refusal rate?","acceptedAnswer":{"@type":"Answer","text":"Abliteration is an editing technique that strips a model's built-in refusals by directly altering its weights, which is only possible because the weights are public. Anthropic's attempt took about 2,200 GPU hours (roughly $4,400) and cut GLM-5.3's refusal rate from over 90% to as low as 2% with little loss of capability; an experienced team could do it for about 600 GPU hours, or $1,200. Anyone assessing open-weight model risk can keep up with techniques like abliteration via daily.dev."}}]} ```