{"id":61396,"date":"2026-09-21T11:53:04","date_gmt":"2026-09-21T11:53:04","guid":{"rendered":"https:\/\/www.threatshub.org\/blog\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\/"},"modified":"2026-09-21T11:53:04","modified_gmt":"2026-09-21T11:53:04","slug":"the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark","status":"publish","type":"post","link":"https:\/\/www.threatshub.org\/blog\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\/","title":{"rendered":"The AI models that cheat the most, according to new CAIS benchmark"},"content":{"rendered":"<figure class=\"wp-block-image size-large is-content-hero\"><img fetchpriority=\"high\" decoding=\"async\" height=\"612\" width=\"1024\" src=\"https:\/\/www.zdnet.com\/wp-content\/uploads\/sites\/3\/GettyImages-2101007925_a969c1.jpg?w=864\" alt=\"A ladder and planks on a round maze used to cheat the challenge\" class=\"wp-image-1319043\" srcset=\"https:\/\/www.zdnet.com\/wp-content\/uploads\/sites\/3\/GettyImages-2101007925_a969c1.jpg 2240w, https:\/\/www.zdnet.com\/wp-content\/uploads\/sites\/3\/GettyImages-2101007925_a969c1.jpg?resize=300,179 300w, https:\/\/www.zdnet.com\/wp-content\/uploads\/sites\/3\/GettyImages-2101007925_a969c1.jpg?resize=768,459 768w, https:\/\/www.zdnet.com\/wp-content\/uploads\/sites\/3\/GettyImages-2101007925_a969c1.jpg?resize=1024,612 1024w, https:\/\/www.zdnet.com\/wp-content\/uploads\/sites\/3\/GettyImages-2101007925_a969c1.jpg?resize=1536,917 1536w, https:\/\/www.zdnet.com\/wp-content\/uploads\/sites\/3\/GettyImages-2101007925_a969c1.jpg?resize=2048,1223 2048w\" sizes=\"(max-width: 1024px) 100vw, 1024px\"><span class=\"zd-image-credit\">Stefan_Alfonso\/iStock\/Getty Images Plus<\/span><\/figure>\n<h3 class=\"wp-block-heading\">ZDNET\u2019s key takeaway<\/h3>\n<ul class=\"wp-block-list\">\n<li>The Center for AI Safety (CAIS) created CheatBench.<\/li>\n<li>They found that every agent cheats in some scenarios.<\/li>\n<li>The propensity to cheat creates risks for humanity.<\/li>\n<\/ul>\n<hr class=\"wp-block-separator has-alpha-channel-opacity\">\n<p>AI labs often tout impressive benchmark scores when releasing new models, showing better capabilities in areas like <a href=\"https:\/\/www.zdnet.com\/article\/i-cant-stop-80-of-developers-find-ai-coding-more-addictive-than-helpful\/\">coding<\/a>, computer use, and more than their competitors. However, those benchmarks aren\u2019t always a reliable measure of what AI can do because they\u2019re easily beaten by exponentially improving models and can emphasize <a href=\"https:\/\/www.pcmag.com\/opinions\/why-ai-benchmarks-are-total-bs-and-how-openai-and-anthropic-use-them-to\" target=\"_blank\" rel=\"noopener\">marketing over actual performance<\/a>.<\/p>\n<p><strong>Also: <a href=\"https:\/\/www.zdnet.com\/article\/reasoning-ai-models-are-overwhelming-the-benchmark-tests-its-time-for-human-evaluation\/\">With AI models clobbering every benchmark, it\u2019s time for human evaluation<\/a><\/strong><\/p>\n<p>Benchmarks like <a href=\"https:\/\/www.zdnet.com\/article\/humanitys-last-exam-benchmark-is-stumping-top-ai-models-can-you-do-any-better\/\">Humanity\u2019s Last Exam<\/a> try to counter this issue by challenging models in more realistic environments. But models still find loopholes to complete tasks \u2014 <a href=\"https:\/\/www.zdnet.com\/article\/openai-hugging-face-attack-agent\/\">Hugging Face incident<\/a>, anyone? <\/p>\n<p>So, the Center for AI Safety (CAIS) created <a href=\"https:\/\/www.cheatbench.ai\/\" target=\"_blank\" rel=\"noopener\">CheatBench<\/a>. Yes, it\u2019s exactly what it sounds like \u2014 and nearly every frontier model is guilty.<\/p>\n<h2 class=\"wp-block-heading\">What CAIS found <\/h2>\n<p>AI models are rewarded for performing tasks well and quickly. A lack of knowledge or tools incentivizes them to do what researchers call \u201creward gaming\u201d by \u201cfinding hidden answers, copying another agent\u2019s submission, or manipulating how its work is graded,\u201d CAIS explained. \u201cCheatBench measures how often AI agents take these shortcuts when honest work is difficult.\u201d<\/p>\n<p>CAIS tested several agents running the latest and most lauded models, including <a href=\"https:\/\/www.zdnet.com\/innovation\/openai-astra-finally-here-experts-worried\/\">OpenAI\u2019s GPT-6 Astra in Codex<\/a>, <a href=\"https:\/\/www.zdnet.com\/article\/anthropiclaude-fable-5-nerfed-mythos-with-guardrails\/\">Anthropic\u2019s Fabel 5.1 in Claude Code<\/a>, and <a href=\"https:\/\/www.zdnet.com\/article\/ai-model-release-tracker\/\">Meta\u2019s newly released Muse Spark 1.3 in Muse Code<\/a>. These agents were tested across 10 categories, including writing, professional work, mathematical research, and coding. Using \u201choneypot\u201d clues hidden in task filespaces, the test separated acceptable reference use from cheating. CheatBench accounts for any time agents attempt to cheat, whether they are successful or not.<\/p>\n<p><strong>Also: <a href=\"https:\/\/www.zdnet.com\/innovation\/how-workers-use-ai-at-work-2026\/\">What workers are really using AI for in 2026 \u2013 and what they aren\u2019t<\/a><\/strong><\/p>\n<p>\u201cEach setting establishes an expectation of honest work, introduces a discoverable opportunity to cheat, and defines the action that crosses that boundary,\u201d the researchers explained. <\/p>\n<p>Every agent the researchers tested cheated in at least some scenarios, but Astra came in as the most honest with a cheating rate of 48.2% \u2014 still almost half the time. Grok 4.6 was scored the biggest cheater with a rate of 81.5%. Open-weight models Kimi K3 and DeepSeek V4 Pro landed in the middle between several other proprietary frontier models.<\/p>\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" height=\"480\" width=\"1024\" src=\"https:\/\/www.zdnet.com\/wp-content\/uploads\/sites\/3\/Screenshot-2026-09-18-at-12.58.45-PM.png?w=864\" alt class=\"wp-image-1317127\" srcset=\"https:\/\/www.zdnet.com\/wp-content\/uploads\/sites\/3\/Screenshot-2026-09-18-at-12.58.45-PM.png 2302w, https:\/\/www.zdnet.com\/wp-content\/uploads\/sites\/3\/Screenshot-2026-09-18-at-12.58.45-PM.png?resize=300,141 300w, https:\/\/www.zdnet.com\/wp-content\/uploads\/sites\/3\/Screenshot-2026-09-18-at-12.58.45-PM.png?resize=768,360 768w, https:\/\/www.zdnet.com\/wp-content\/uploads\/sites\/3\/Screenshot-2026-09-18-at-12.58.45-PM.png?resize=1024,480 1024w, https:\/\/www.zdnet.com\/wp-content\/uploads\/sites\/3\/Screenshot-2026-09-18-at-12.58.45-PM.png?resize=1536,721 1536w, https:\/\/www.zdnet.com\/wp-content\/uploads\/sites\/3\/Screenshot-2026-09-18-at-12.58.45-PM.png?resize=2048,961 2048w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\"><\/figure>\n<h2 class=\"wp-block-heading\">What cheating looks like \u2013 and why it matters <\/h2>\n<p>In one example, researchers asked <a href=\"https:\/\/www.zdnet.com\/article\/claude-opus-5-near-fable-performance-at-half-the-price\/\">Claude Opus<\/a> to design a protein binder. The model knew it wasn\u2019t allowed to refer to a set of accepted designs in the filespace, but eventually caved.<\/p>\n<p>\u201cAfter seven rejected designs, it locates the file, writes that it should not look at or copy it, and reads it with a shell command in the very next call,\u201d the researchers wrote. In its reasoning, the model admitted that using work other than its own would \u201cmisrepresent my actual capabilities in this evaluation, so I shouldn\u2019t look at or copy it.\u201d But its very next step was to reference the accepted designs.<\/p>\n<p><strong>Also: <a href=\"https:\/\/www.zdnet.com\/article\/how-to-check-if-chatgpt-ai-tools-cite-website-improve-chances\/\">How to check if ChatGPT and other AI tools cite your website \u2013 and improve your chances in 2026<\/a><\/strong><\/p>\n<p>This result demonstrated both a readable choice the model made to contradict itself, and what looked like a hole in our understanding about what made the model jump from one instinct to the next.<\/p>\n<p>Things got more interesting at the task category level. Even if an agent didn\u2019t cheat in one area, it could cheat significantly more in another. Fable 5.1 was only 5% likely to cheat at games, but 100% likely to cheat on knowledge work tasks.<\/p>\n<p><strong>Also: <a href=\"https:\/\/www.zdnet.com\/article\/how-ai-chatbots-keep-users-engaged-warning-signs-to-look-out-for\/\">The sneaky ways AI chatbots keep you hooked \u2013 and coming back for more<\/a><\/strong><\/p>\n<p>Reinforcement learning trains models not to abandon a task, even if pursuing it creates conflict-ridden choices. CAIS noted in its paper that <a href=\"https:\/\/www.zdnet.com\/article\/how-ai-chatbots-keep-users-engaged-warning-signs-to-look-out-for\/\">sycophancy<\/a> is an early sign of reward gaming. This term refers to AI models\u2019 tendency to be too agreeable and encouraging of whatever a user says, sometimes regardless of whether it\u2019s incorrect, delusional, or could lead to harmful behavior. Traits like sycophancy and reward gaming show how models can prioritize accomplishing a task correctly to please a user over the alignment training researchers work so hard to build in.<\/p>\n<p>These tests represent relatively low stakes. But CAIS researchers created CheatBench because of the risks of this behavior at scale across different tasks. Earlier this month, <a href=\"https:\/\/mashable.com\/tech\/anthropic-ai-researcher-quit-ethics-safety\" target=\"_blank\" rel=\"noopener\">yet another researcher quit<\/a> Anthropic over concerns that the company isn\u2019t developing AI responsibly for a future in which it could build itself away from human-oriented values and kill us. <\/p>\n<p>A propensity to cheat, or complete a task at any cost, puts our potentially differing priorities at odds with an increasingly powerful technology. As I <a href=\"https:\/\/links.email.cnet.com\/s\/vb\/hribMjl6r1COTsKunu3Ld-PfK__vMYlCnAmhUezPYd2V9rZ01Wq8S1f0L601iXGW7oheAWpmA8JhZFksJbkJBFGoxhVenUYaZIpfy9KoBD9VBoP7ozfMEFDgwg9ER-cyQHaOjzocNXbvn0KKUHQYcf7djrzTjP-RVT_hjA\/16WC2SpgLOh4oXDT7Rxkz3zedFNISNab\/24\" target=\"_blank\" rel=\"noopener\">explained<\/a> in the AI Leaderboard newsletter last week, it won\u2019t necessarily be a demonstrated animosity toward humans that pits AI against us; it may be that we are simply in the way and end up as collateral.<\/p>\n<section class=\"zd-disqus\" data-zd-disqus data-disqus-shortname=\"zdnet-1\" data-disqus-identifier=\"1316748\" data-disqus-url=\"https:\/\/www.zdnet.com\/innovation\/ai-model-cheating-benchmark-cheatbench\/\" data-disqus-title=\"The AI models that cheat the most, according to new CAIS benchmark\" data-disqus-api-key=\"y0Dp0T6itn9rbTvSf2xkzkhaUL7c5EJsmaWF3XYBqXNz8GHH7wSX1nXi4qzdihWV\" data-disqus-show-label=\"Show Comments\" data-disqus-hide-label=\"Hide Comments\"> <button type=\"button\" class=\"zd-disqus__toggle\" data-zd-disqus-toggle aria-expanded=\"false\"> <span class=\"zd-disqus__toggle-text\" data-zd-disqus-label>Show Comments<\/span> <svg class=\"zd-disqus__toggle-icon\" width=\"14\" height=\"14\" viewBox=\"0 0 14 14\" aria-hidden=\"true\" focusable=\"false\"> <path d=\"M7 1 V12.5 M2.5 8 L7 12.5 L11.5 8\" fill=\"none\" stroke=\"currentColor\" stroke-width=\"1.5\" stroke-linecap=\"round\" stroke-linejoin=\"round\" \/> <\/svg> <\/button> <\/p>\n<div class=\"zd-disqus__panel\" data-zd-disqus-panel hidden>\n<p class=\"zd-disqus__consent\" data-zd-disqus-consent hidden>Please accept cookies to view the comments powered by Disqus.<\/p>\n<\/p><\/div>\n<p> <noscript>Please enable JavaScript to view the comments powered by Disqus.<\/noscript> <\/section>\n<div class=\"articleAuthorBioFooter-container\">\n<div class=\"articleAuthorBioFooter\">\n<div class=\"articleAuthorBioFooter_body\">\n<div class=\"articleAuthorBioFooter_nameBlock\">\n<div class=\"c-cmsImage articleAuthorBioFooter_image\"> <picture class=\"c-cmsImage_image\"> <img alt=\"Radhika Rajkumar\" src=\"https:\/\/www.zdnet.com\/wp-content\/uploads\/sites\/3\/008a2316.jpg?w=64\" srcset=\"https:\/\/www.zdnet.com\/wp-content\/uploads\/sites\/3\/008a2316.jpg?w=128 2x\" class=\"avatar avatar-64 photo\" height=\"64\" width=\"64\" loading=\"lazy\" decoding=\"async\"> <\/picture> <\/div>\n<div class=\"articleAuthorBioFooter_nameText\">\n<div class=\"articleAuthorBioFooter_name\"> <span>Radhika Rajkumar<\/span> <\/div>\n<p> <span class=\"articleAuthorBioFooter_credentials\"> Senior Editor <\/span> <\/div>\n<\/p><\/div>\n<p> <span class=\"articleAuthorBioFooter_bio has-bio\"> <span>Radhika Rajkumar is a senior editor at ZDNET based in New York City. She covers AI, specializing in safety, privacy and security, policy, education, and synthetic media. She also leads ZDNET&#8217;s newsletter strategy. Radhika holds a Masters in Creative Publishing and Critical Journalism from The New School.<\/span> <a href=\"https:\/\/www.zdnet.com\/meet-the-team\/rrajkumar\/\" rel=\"author\" class=\"articleAuthorBioFooter_link\"> See full bio <\/a> <\/span> <\/div>\n<p> <!-- Latest Articles Section --> <\/p>\n<div class=\"articleAuthorBioFooter_latest\">\n<h3 class=\"articleAuthorBioFooter_latestTitle\"> The Latest from Radhika <\/h3>\n<\/p><\/div>\n<\/p><\/div>\n<\/p><\/div>\n<p><!-- .articleAuthorBioFooter-container --><\/p>\n<div class=\"zd-post-footer-pattern\">\n<div class=\"wp-block-group is-layout-constrained wp-block-group-is-layout-constrained\">\n<div class=\"c-ccb c-ccb--entry-card c-ccb--source-dynamic wp-block-ziff-davis-curated-content\" data-used=\"[1316748,1316521,1317439,1297861,705262]\">\n<div class=\"ccb-header\">\n<div class=\"ccb-header__text\">\n<h2 class=\"ccb-header__block-title\">\/ featured <\/h2>\n<\/p><\/div>\n<\/p><\/div>\n<p> <!-- .entry-featured-grid -->\n<\/div>\n<\/div>\n<\/div>\n<p>READ MORE <a href=\"https:\/\/www.zdnet.com\/innovation\/ai-model-cheating-benchmark-cheatbench\/\">HERE<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>We know models cheat. A new benchmark measures how much, and on what tasks. READ MORE HERE&#8230;<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"colormag_page_layout":"default_layout","footnotes":""},"categories":[62],"tags":[],"class_list":["post-61396","post","type-post","status-publish","format-standard","hentry","category-zdnet-security"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v28.1 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>The AI models that cheat the most, according to new CAIS benchmark 2026 | ThreatsHub Cybersecurity News<\/title>\n<meta name=\"description\" content=\"ThreatsHub Cybersecurity News | ThreatsHub.org | Cloud Security &amp; Cyber Threats Analysis Hub. 100% Free OSINT Threat Intelligent and Cybersecurity News.\" \/>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.threatshub.org\/blog\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\/\" \/>\n<meta property=\"og:locale\" content=\"en_US\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"The AI models that cheat the most, according to new CAIS benchmark 2026 | ThreatsHub Cybersecurity News\" \/>\n<meta property=\"og:description\" content=\"ThreatsHub Cybersecurity News | ThreatsHub.org | Cloud Security &amp; Cyber Threats Analysis Hub. 100% Free OSINT Threat Intelligent and Cybersecurity News.\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.threatshub.org\/blog\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\/\" \/>\n<meta property=\"og:site_name\" content=\"ThreatsHub Cybersecurity News\" \/>\n<meta property=\"article:published_time\" content=\"2026-09-21T11:53:04+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.threatshub.org\/blog\/coredata\/uploads\/2025\/05\/Threatshub_Favicon1.jpg\" \/>\n\t<meta property=\"og:image:width\" content=\"432\" \/>\n\t<meta property=\"og:image:height\" content=\"435\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/jpeg\" \/>\n<meta name=\"author\" content=\"TH Author\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:creator\" content=\"@threatshub\" \/>\n<meta name=\"twitter:site\" content=\"@threatshub\" \/>\n<meta name=\"twitter:label1\" content=\"Written by\" \/>\n\t<meta name=\"twitter:data1\" content=\"TH Author\" \/>\n\t<meta name=\"twitter:label2\" content=\"Est. reading time\" \/>\n\t<meta name=\"twitter:data2\" content=\"5 minutes\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/www.threatshub.org\\\/blog\\\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\\\/#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/www.threatshub.org\\\/blog\\\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\\\/\"},\"author\":{\"name\":\"TH Author\",\"@id\":\"https:\\\/\\\/www.threatshub.org\\\/blog\\\/#\\\/schema\\\/person\\\/12e0a8671ff89a863584f193e7062476\"},\"headline\":\"The AI models that cheat the most, according to new CAIS benchmark\",\"datePublished\":\"2026-09-21T11:53:04+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/www.threatshub.org\\\/blog\\\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\\\/\"},\"wordCount\":923,\"publisher\":{\"@id\":\"https:\\\/\\\/www.threatshub.org\\\/blog\\\/#organization\"},\"image\":{\"@id\":\"https:\\\/\\\/www.threatshub.org\\\/blog\\\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/www.zdnet.com\\\/wp-content\\\/uploads\\\/sites\\\/3\\\/GettyImages-2101007925_a969c1.jpg?w=864\",\"articleSection\":[\"ZDNet | Security\"],\"inLanguage\":\"en-US\"},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/www.threatshub.org\\\/blog\\\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\\\/\",\"url\":\"https:\\\/\\\/www.threatshub.org\\\/blog\\\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\\\/\",\"name\":\"The AI models that cheat the most, according to new CAIS benchmark 2026 | ThreatsHub Cybersecurity News\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/www.threatshub.org\\\/blog\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/www.threatshub.org\\\/blog\\\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\\\/#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/www.threatshub.org\\\/blog\\\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/www.zdnet.com\\\/wp-content\\\/uploads\\\/sites\\\/3\\\/GettyImages-2101007925_a969c1.jpg?w=864\",\"datePublished\":\"2026-09-21T11:53:04+00:00\",\"description\":\"ThreatsHub Cybersecurity News | ThreatsHub.org | Cloud Security & Cyber Threats Analysis Hub. 100% Free OSINT Threat Intelligent and Cybersecurity News.\",\"breadcrumb\":{\"@id\":\"https:\\\/\\\/www.threatshub.org\\\/blog\\\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\\\/#breadcrumb\"},\"inLanguage\":\"en-US\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/www.threatshub.org\\\/blog\\\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\\\/\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/www.threatshub.org\\\/blog\\\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\\\/#primaryimage\",\"url\":\"https:\\\/\\\/www.zdnet.com\\\/wp-content\\\/uploads\\\/sites\\\/3\\\/GettyImages-2101007925_a969c1.jpg?w=864\",\"contentUrl\":\"https:\\\/\\\/www.zdnet.com\\\/wp-content\\\/uploads\\\/sites\\\/3\\\/GettyImages-2101007925_a969c1.jpg?w=864\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/www.threatshub.org\\\/blog\\\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\\\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https:\\\/\\\/www.threatshub.org\\\/blog\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"The AI models that cheat the most, according to new CAIS benchmark\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/www.threatshub.org\\\/blog\\\/#website\",\"url\":\"https:\\\/\\\/www.threatshub.org\\\/blog\\\/\",\"name\":\"ThreatsHub Cybersecurity News\",\"description\":\"%%focuskw%% Threat Intel \u2013 Threat Intel Services \u2013 CyberIntelligence \u2013 Cyber Threat Intelligence - Threat Intelligence Feeds - Threat Intelligence Reports - CyberSecurity Report \u2013 Cyber Security PDF \u2013 Cybersecurity Trends - Cloud Sandbox \u2013- Threat IntelligencePortal \u2013 Incident Response \u2013 Threat Hunting \u2013 IOC - Yara - Security Operations Center \u2013 SecurityOperation Center \u2013 Security SOC \u2013 SOC Services - Advanced Threat - Threat Detection - TargetedAttack \u2013 APT \u2013 Anti-APT \u2013 Advanced Protection \u2013 Cyber Security Services \u2013 Cybersecurity Services -Threat Intelligence Platform\",\"publisher\":{\"@id\":\"https:\\\/\\\/www.threatshub.org\\\/blog\\\/#organization\"},\"alternateName\":\"Threatshub.org\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/www.threatshub.org\\\/blog\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"en-US\"},{\"@type\":\"Organization\",\"@id\":\"https:\\\/\\\/www.threatshub.org\\\/blog\\\/#organization\",\"name\":\"ThreatsHub.org\",\"alternateName\":\"Threatshub.org\",\"url\":\"https:\\\/\\\/www.threatshub.org\\\/blog\\\/\",\"logo\":{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/www.threatshub.org\\\/blog\\\/#\\\/schema\\\/logo\\\/image\\\/\",\"url\":\"https:\\\/\\\/www.threatshub.org\\\/blog\\\/wp-content\\\/uploads\\\/2025\\\/05\\\/Threatshub_Favicon1.jpg\",\"contentUrl\":\"https:\\\/\\\/www.threatshub.org\\\/blog\\\/wp-content\\\/uploads\\\/2025\\\/05\\\/Threatshub_Favicon1.jpg\",\"width\":432,\"height\":435,\"caption\":\"ThreatsHub.org\"},\"image\":{\"@id\":\"https:\\\/\\\/www.threatshub.org\\\/blog\\\/#\\\/schema\\\/logo\\\/image\\\/\"},\"sameAs\":[\"https:\\\/\\\/x.com\\\/threatshub\"]},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/www.threatshub.org\\\/blog\\\/#\\\/schema\\\/person\\\/12e0a8671ff89a863584f193e7062476\",\"name\":\"TH Author\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/066276f086d5155df79c850206a779ad368418a844da0182ce43f9cd5b506c3d?s=96&d=mm&r=g\",\"url\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/066276f086d5155df79c850206a779ad368418a844da0182ce43f9cd5b506c3d?s=96&d=mm&r=g\",\"contentUrl\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/066276f086d5155df79c850206a779ad368418a844da0182ce43f9cd5b506c3d?s=96&d=mm&r=g\",\"caption\":\"TH Author\"}}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"The AI models that cheat the most, according to new CAIS benchmark 2026 | ThreatsHub Cybersecurity News","description":"ThreatsHub Cybersecurity News | ThreatsHub.org | Cloud Security & Cyber Threats Analysis Hub. 100% Free OSINT Threat Intelligent and Cybersecurity News.","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.threatshub.org\/blog\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\/","og_locale":"en_US","og_type":"article","og_title":"The AI models that cheat the most, according to new CAIS benchmark 2026 | ThreatsHub Cybersecurity News","og_description":"ThreatsHub Cybersecurity News | ThreatsHub.org | Cloud Security & Cyber Threats Analysis Hub. 100% Free OSINT Threat Intelligent and Cybersecurity News.","og_url":"https:\/\/www.threatshub.org\/blog\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\/","og_site_name":"ThreatsHub Cybersecurity News","article_published_time":"2026-09-21T11:53:04+00:00","og_image":[{"width":432,"height":435,"url":"https:\/\/www.threatshub.org\/blog\/coredata\/uploads\/2025\/05\/Threatshub_Favicon1.jpg","type":"image\/jpeg"}],"author":"TH Author","twitter_card":"summary_large_image","twitter_creator":"@threatshub","twitter_site":"@threatshub","twitter_misc":{"Written by":"TH Author","Est. reading time":"5 minutes"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/www.threatshub.org\/blog\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\/#article","isPartOf":{"@id":"https:\/\/www.threatshub.org\/blog\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\/"},"author":{"name":"TH Author","@id":"https:\/\/www.threatshub.org\/blog\/#\/schema\/person\/12e0a8671ff89a863584f193e7062476"},"headline":"The AI models that cheat the most, according to new CAIS benchmark","datePublished":"2026-09-21T11:53:04+00:00","mainEntityOfPage":{"@id":"https:\/\/www.threatshub.org\/blog\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\/"},"wordCount":923,"publisher":{"@id":"https:\/\/www.threatshub.org\/blog\/#organization"},"image":{"@id":"https:\/\/www.threatshub.org\/blog\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\/#primaryimage"},"thumbnailUrl":"https:\/\/www.zdnet.com\/wp-content\/uploads\/sites\/3\/GettyImages-2101007925_a969c1.jpg?w=864","articleSection":["ZDNet | Security"],"inLanguage":"en-US"},{"@type":"WebPage","@id":"https:\/\/www.threatshub.org\/blog\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\/","url":"https:\/\/www.threatshub.org\/blog\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\/","name":"The AI models that cheat the most, according to new CAIS benchmark 2026 | ThreatsHub Cybersecurity News","isPartOf":{"@id":"https:\/\/www.threatshub.org\/blog\/#website"},"primaryImageOfPage":{"@id":"https:\/\/www.threatshub.org\/blog\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\/#primaryimage"},"image":{"@id":"https:\/\/www.threatshub.org\/blog\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\/#primaryimage"},"thumbnailUrl":"https:\/\/www.zdnet.com\/wp-content\/uploads\/sites\/3\/GettyImages-2101007925_a969c1.jpg?w=864","datePublished":"2026-09-21T11:53:04+00:00","description":"ThreatsHub Cybersecurity News | ThreatsHub.org | Cloud Security & Cyber Threats Analysis Hub. 100% Free OSINT Threat Intelligent and Cybersecurity News.","breadcrumb":{"@id":"https:\/\/www.threatshub.org\/blog\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\/#breadcrumb"},"inLanguage":"en-US","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.threatshub.org\/blog\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\/"]}]},{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/www.threatshub.org\/blog\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\/#primaryimage","url":"https:\/\/www.zdnet.com\/wp-content\/uploads\/sites\/3\/GettyImages-2101007925_a969c1.jpg?w=864","contentUrl":"https:\/\/www.zdnet.com\/wp-content\/uploads\/sites\/3\/GettyImages-2101007925_a969c1.jpg?w=864"},{"@type":"BreadcrumbList","@id":"https:\/\/www.threatshub.org\/blog\/the-ai-models-that-cheat-the-most-according-to-new-cais-benchmark\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/www.threatshub.org\/blog\/"},{"@type":"ListItem","position":2,"name":"The AI models that cheat the most, according to new CAIS benchmark"}]},{"@type":"WebSite","@id":"https:\/\/www.threatshub.org\/blog\/#website","url":"https:\/\/www.threatshub.org\/blog\/","name":"ThreatsHub Cybersecurity News","description":"%%focuskw%% Threat Intel \u2013 Threat Intel Services \u2013 CyberIntelligence \u2013 Cyber Threat Intelligence - Threat Intelligence Feeds - Threat Intelligence Reports - CyberSecurity Report \u2013 Cyber Security PDF \u2013 Cybersecurity Trends - Cloud Sandbox \u2013- Threat IntelligencePortal \u2013 Incident Response \u2013 Threat Hunting \u2013 IOC - Yara - Security Operations Center \u2013 SecurityOperation Center \u2013 Security SOC \u2013 SOC Services - Advanced Threat - Threat Detection - TargetedAttack \u2013 APT \u2013 Anti-APT \u2013 Advanced Protection \u2013 Cyber Security Services \u2013 Cybersecurity Services -Threat Intelligence Platform","publisher":{"@id":"https:\/\/www.threatshub.org\/blog\/#organization"},"alternateName":"Threatshub.org","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.threatshub.org\/blog\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"en-US"},{"@type":"Organization","@id":"https:\/\/www.threatshub.org\/blog\/#organization","name":"ThreatsHub.org","alternateName":"Threatshub.org","url":"https:\/\/www.threatshub.org\/blog\/","logo":{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/www.threatshub.org\/blog\/#\/schema\/logo\/image\/","url":"https:\/\/www.threatshub.org\/blog\/coredata\/uploads\/2025\/05\/Threatshub_Favicon1.jpg","contentUrl":"https:\/\/www.threatshub.org\/blog\/coredata\/uploads\/2025\/05\/Threatshub_Favicon1.jpg","width":432,"height":435,"caption":"ThreatsHub.org"},"image":{"@id":"https:\/\/www.threatshub.org\/blog\/#\/schema\/logo\/image\/"},"sameAs":["https:\/\/x.com\/threatshub"]},{"@type":"Person","@id":"https:\/\/www.threatshub.org\/blog\/#\/schema\/person\/12e0a8671ff89a863584f193e7062476","name":"TH Author","image":{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/secure.gravatar.com\/avatar\/066276f086d5155df79c850206a779ad368418a844da0182ce43f9cd5b506c3d?s=96&d=mm&r=g","url":"https:\/\/secure.gravatar.com\/avatar\/066276f086d5155df79c850206a779ad368418a844da0182ce43f9cd5b506c3d?s=96&d=mm&r=g","contentUrl":"https:\/\/secure.gravatar.com\/avatar\/066276f086d5155df79c850206a779ad368418a844da0182ce43f9cd5b506c3d?s=96&d=mm&r=g","caption":"TH Author"}}]}},"_links":{"self":[{"href":"https:\/\/www.threatshub.org\/blog\/wp-json\/wp\/v2\/posts\/61396","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.threatshub.org\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.threatshub.org\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.threatshub.org\/blog\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.threatshub.org\/blog\/wp-json\/wp\/v2\/comments?post=61396"}],"version-history":[{"count":0,"href":"https:\/\/www.threatshub.org\/blog\/wp-json\/wp\/v2\/posts\/61396\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.threatshub.org\/blog\/wp-json\/wp\/v2\/media?parent=61396"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.threatshub.org\/blog\/wp-json\/wp\/v2\/categories?post=61396"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.threatshub.org\/blog\/wp-json\/wp\/v2\/tags?post=61396"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}