--- title: Booz Allen's AI cyber threat index: Claude Mythos tops... description: Booz Allen Hamilton's first Cyber Weapon Index tested 18 AI models — nine American, nine Chinese — against a live corporate network to gauge autonomous... canonical: https://daily.dev/posts/booz-allen-s-ai-cyber-threat-index-claude-mythos-tops-the-ranking-but-the-harness-matters-more-tha-uwzrutoqu twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Booz Allen's AI cyber threat index: Claude Mythos tops the ranking, but the harness matters more than the model | daily.dev og:description: Booz Allen Hamilton's first Cyber Weapon Index tested 18 AI models — nine American, nine Chinese — against a live corporate network to gauge autonomous... og:url: https://daily.dev/posts/booz-allen-s-ai-cyber-threat-index-claude-mythos-tops-the-ranking-but-the-harness-matters-more-tha-uwzrutoqu og:image: https://api.daily.dev/og/posts/UwzrUTOqU.png og:image:alt: Booz Allen's AI cyber threat index: Claude Mythos tops the ranking, but the harness matters more than the model 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. # Booz Allen's AI cyber threat index: Claude Mythos tops the ranking, but the harness matters more than the model **[Collections](https://daily.dev/sources/collections)** · 3 min read · 4 upvotes · 0 comments ## Summary Booz Allen Hamilton's first Cyber Weapon Index tested 18 AI models — nine American, nine Chinese — against a live corporate network to gauge autonomous cyberattack capability. Anthropic's Claude Mythos topped the ranking with a score of 80, the only model to complete a full cyber kill chain with and without stolen credentials. Against real-world production vulnerabilities, all nine frontier models scored zero, with one partial exception credited to Mythos. The more consequential finding: pairing a low-ranked model (Claude Sonnet 5, ranked 15th) with a purpose-built attack harness closed 67 of the 80-point gap to Mythos, suggesting the harness and surrounding system — not the base model — is the real unit of risk. Guardrails also behaved inconsistently across sibling models on identical tasks. OpenAI's Astra was excluded without explanation despite OpenAI's own claims about crossing a cybersecurity threshold. Booz Allen is calling for regulatory deadlines and a national testing program, while also promoting its own commercial counter-AI product alongside the report. ## Content # Booz Allen's AI cyber benchmark finds attack harnesses matter more than raw model capability Booz Allen Hamilton published its first Cyber Weapon Index this week, scoring 18 AI models — nine American, nine Chinese — on their ability to autonomously conduct offensive cyber operations against a live corporate network. The headline result is striking, but the more interesting finding quietly undermines it. ## What the benchmark tested Models were evaluated on unassisted network intrusion and vulnerability research: no human guidance, no hand-holding, just the model and a production-grade target. The scoring ran from 0 to 100 based on how far each model progressed through the kill chain. Anthropic's Claude Mythos came out on top with a score of 80 and was the only model to complete the full kill chain autonomously — achieving domain administrator access both with stolen credentials and without them. Grok-4.5 from SpaceXAI scored 49, and several others including Muse Spark 1.1 and GLM-5.2 reached full domain access but couldn't complete the chain unassisted. Alibaba's Qwen3-Coder finished last with a score of 4. Against real-world, previously unseen vulnerabilities in production code, all nine frontier models tested scored zero — though the report also credits Mythos elsewhere with understanding and exploiting such a flaw, which is a contradiction the report doesn't fully resolve. ## The finding that undercuts the ranking Here's where it gets interesting. Booz Allen paired Claude Sonnet 5 — ranked 15th with a score of 13 — with an attack harness, a software scaffold that structures the model's actions and manages its workflow. That combination closed 67 points of the gap to Mythos. That's a cheap piece of software nearly erasing a 67-point difference between models. Booz Allen's own conclusion: the harness and surrounding system, not the raw model capability, is the real unit of risk. A mediocre model with good scaffolding can punch well above its weight. The report also found that guardrails were inconsistent across sibling models given identical tasks — meaning safety behavior isn't reliably inherited even within the same model family. ## The six-month warning Booz Allen warns that most tested models could reach Mythos' weaponization level within six months, and calls mainstream AI-driven cyberattacks "imminent." The UK AI Security Institute and OpenAI have reported similar findings in capture-the-flag tests, though with lower success rates. The concern isn't just frontier models. Open-weight models — which anyone can download and run — will soon make autonomous attacks cheap enough to scale broadly. At that point, human-speed defense response times won't be sufficient. Notably, OpenAI's Astra model was excluded from testing entirely, despite OpenAI stating it had crossed a critical cybersecurity threshold. That's a gap worth flagging. ## Proposed responses Booz Allen published the index alongside its own commercial counter-AI product, Vellox Labs Guile, a deception system that reportedly fools AI attackers more than 90% of the time — though it rarely works on human attackers, which says something about how differently AI agents navigate networks. The report calls for regulatory deadlines and a national testing program covering foreign and open-weight models. The broader argument from security researchers is that defense pipelines need to become automated too, since human analysts can't match the speed of autonomous attack chains. The benchmark is a useful data point, but the harness finding is the one that should stick. The question isn't just which model scores highest in isolation — it's what happens when someone wraps a middling model in good tooling and points it at your network. ## Questions this post answers ### Which AI model scored highest on Booz Allen's Cyber Weapon Index for autonomous cyberattacks? Anthropic's Claude Mythos topped Booz Allen Hamilton's Cyber Weapon Index with a score of 80 out of 100, the only model among 18 tested (nine American, nine Chinese) to autonomously complete the full attack kill chain, achieving domain administrator access both with and without stolen credentials. Grok-4.5 scored 49, while Qwen3-Coder finished last at 4. _Anyone tracking AI cyber risk benchmarks can follow model and tooling comparisons like this on daily.dev._ ### Does a good attack harness matter more than the underlying AI model for autonomous hacking? Yes, according to Booz Allen Hamilton's testing, which paired Claude Sonnet 5 (ranked 15th with a score of 13) with an attack harness — a software scaffold structuring the model's actions — and closed 67 of the 80-point gap to the top-scoring Claude Mythos. This suggests a mediocre model with strong scaffolding can approach top-tier attack performance. _Teams assessing AI security risk can weigh tooling and harness design alongside model choice via daily.dev._ ### Why was OpenAI's Astra model excluded from Booz Allen's Cyber Weapon Index testing? OpenAI's Astra model was excluded from testing entirely despite OpenAI itself stating the model had crossed a critical cybersecurity threshold, a gap Booz Allen Hamilton's report does not explain. This omission is notable because Astra's exclusion leaves an open question about how it would compare to Claude Mythos, the top-scoring model in the index. _Security-minded developers evaluating frontier model risk profiles can track benchmark gaps like this on daily.dev._ ## Similar posts on daily.dev - [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 - [Claude Mythos Preview completes full cyberattack simulation for the first time](https://daily.dev/posts/claude-mythos-preview-completes-full-cyberattack-simulation-for-the-first-time-dbsyc6gmb) · The New Stack · 0 upvotes · 0 comments - [Claude Mythos Preview Exposes Hidden Code Flaws Fast](https://daily.dev/posts/claude-mythos-preview-exposes-hidden-code-flaws-fast-wdwyjdo3x) · IEEE Spectrum · 0 upvotes · 0 comments --- Tags: [#ai](https://daily.dev/tags/ai), [#cyber](https://daily.dev/tags/cyber), [#claude](https://daily.dev/tags/claude), [#anthropic](https://daily.dev/tags/anthropic), [#ai-security](https://daily.dev/tags/ai-security) [View this post on daily.dev](https://daily.dev/posts/booz-allen-s-ai-cyber-threat-index-claude-mythos-tops-the-ranking-but-the-harness-matters-more-tha-uwzrutoqu) ```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":"Booz Allen's AI cyber threat index: Claude Mythos tops the ranking, but the harness matters more than the model","url":"https://daily.dev/posts/booz-allen-s-ai-cyber-threat-index-claude-mythos-tops-the-ranking-but-the-harness-matters-more-tha-uwzrutoqu","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/booz-allen-s-ai-cyber-threat-index-claude-mythos-tops-the-ranking-but-the-harness-matters-more-tha-uwzrutoqu"},"datePublished":"2026-09-03T09:15:48.529Z","dateModified":"2026-09-13T20:05:09.388Z","description":"Booz Allen Hamilton's first Cyber Weapon Index tested 18 AI models — nine American, nine Chinese — against a live corporate network to gauge autonomous...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/90ae77398aa9763754fde92bfa1eab29?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/90ae77398aa9763754fde92bfa1eab29?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Collections","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":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/booz-allen-s-ai-cyber-threat-index-claude-mythos-tops-the-ranking-but-the-harness-matters-more-tha-uwzrutoqu","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":4},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai,cyber,claude,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":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Booz Allen's AI cyber threat index: Claude Mythos tops the ranking, but the harness matters more than the model"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/booz-allen-s-ai-cyber-threat-index-claude-mythos-tops-the-ranking-but-the-harness-matters-more-tha-uwzrutoqu#faq","mainEntity":[{"@type":"Question","name":"Which AI model scored highest on Booz Allen's Cyber Weapon Index for autonomous cyberattacks?","acceptedAnswer":{"@type":"Answer","text":"Anthropic's Claude Mythos topped Booz Allen Hamilton's Cyber Weapon Index with a score of 80 out of 100, the only model among 18 tested (nine American, nine Chinese) to autonomously complete the full attack kill chain, achieving domain administrator access both with and without stolen credentials. Grok-4.5 scored 49, while Qwen3-Coder finished last at 4. Anyone tracking AI cyber risk benchmarks can follow model and tooling comparisons like this on daily.dev."}},{"@type":"Question","name":"Does a good attack harness matter more than the underlying AI model for autonomous hacking?","acceptedAnswer":{"@type":"Answer","text":"Yes, according to Booz Allen Hamilton's testing, which paired Claude Sonnet 5 (ranked 15th with a score of 13) with an attack harness — a software scaffold structuring the model's actions — and closed 67 of the 80-point gap to the top-scoring Claude Mythos. This suggests a mediocre model with strong scaffolding can approach top-tier attack performance. Teams assessing AI security risk can weigh tooling and harness design alongside model choice via daily.dev."}},{"@type":"Question","name":"Why was OpenAI's Astra model excluded from Booz Allen's Cyber Weapon Index testing?","acceptedAnswer":{"@type":"Answer","text":"OpenAI's Astra model was excluded from testing entirely despite OpenAI itself stating the model had crossed a critical cybersecurity threshold, a gap Booz Allen Hamilton's report does not explain. This omission is notable because Astra's exclusion leaves an open question about how it would compare to Claude Mythos, the top-scoring model in the index. Security-minded developers evaluating frontier model risk profiles can track benchmark gaps like this on daily.dev."}}]} ```