[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-mythos-anthropic-unreleased-ai-model-explained-en":3,"tags-mythos-anthropic-unreleased-ai-model-explained-en":29,"related-lang-mythos-anthropic-unreleased-ai-model-explained-en":30,"related-posts-mythos-anthropic-unreleased-ai-model-explained-en":34,"series-research-fd36cdcc-d9b7-4d57-b64d-f89c8ad531a5":71},{"id":4,"title":5,"content":6,"summary":7,"source":8,"source_url":9,"author":10,"image_url":11,"keywords":12,"language":18,"translated_content":10,"views":19,"is_premium":20,"created_at":21,"updated_at":21,"cover_image":11,"published_at":22,"rewrite_status":23,"rewrite_error":10,"rewritten_from_id":24,"slug":25,"category":26,"related_article_id":27,"status":28,"google_indexed_at":10,"x_posted_at":10},"fd36cdcc-d9b7-4d57-b64d-f89c8ad531a5","Mythos, Anthropic’s unreleased AI model, explained","\u003Cp>\u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\" target=\"_blank\" rel=\"noopener\">Anthropic\u003C\u002Fa> says its new model, \u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002Fnews\" target=\"_blank\" rel=\"noopener\">Mythos\u003C\u002Fa>, is too dangerous for a public release. That is a rare call in AI, and it lands with real weight: the company says the model can beat older systems by 31 percentage points on the USAMO 2026 math benchmark and can complete expert-level hacking tasks 73% of the time in outside testing.\u003C\u002Fp>\u003Cp>The bigger story is not the model itself. It is the choice to keep it behind a restricted rollout while a small set of organizations use it for defensive cyber\u003Ca href=\"\u002Fnews\u002Fopenai-gpt-54-cyber-security-access-en\">security work\u003C\u002Fa>. That decision has triggered a familiar split in AI policy: some see a serious warning sign, while others think the reaction is running ahead of the evidence.\u003C\u002Fp>\u003Ch2>What Mythos actually is\u003C\u002Fh2>\u003Cp>Mythos is Anthropic’s newest large language model, and the company describes it as unusually strong at software engineering, bug finding, and mathematical reasoning. In the 245-page technical document released with the announcement, Anthropic says the model behaves like a senior software engineer in many coding tasks, including spotting subtle errors and correcting its own mistakes.\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1776738631321-l0a3.png\" alt=\"Mythos, Anthropic’s unreleased AI model, explained\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>That matters because the same skills that help a model write cleaner code also help it break software. A system that can reason through a tricky codebase, identify weak points, and adapt after a failed attempt is useful for defenders. It is also useful for attackers who want to turn a bug into an exploit.\u003C\u002Fp>\u003Cp>Anthropic has not opened Mythos to the public. Instead, it is giving limited access through \u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002Fnews\u002Fproject-glasswing\" target=\"_blank\" rel=\"noopener\">Project Glasswing\u003C\u002Fa>, a program meant for cybersecurity testing and patching. The first group includes \u003Ca href=\"https:\u002F\u002Fwww.microsoft.com\" target=\"_blank\" rel=\"noopener\">Microsoft\u003C\u002Fa>, \u003Ca href=\"https:\u002F\u002Fwww.google.com\" target=\"_blank\" rel=\"noopener\">Google\u003C\u002Fa>, \u003Ca href=\"https:\u002F\u002Fwww.apple.com\" target=\"_blank\" rel=\"noopener\">Apple\u003C\u002Fa>, \u003Ca href=\"https:\u002F\u002Faws.amazon.com\" target=\"_blank\" rel=\"noopener\">Amazon Web Services\u003C\u002Fa>, \u003Ca href=\"https:\u002F\u002Fwww.jpmorganchase.com\" target=\"_blank\" rel=\"noopener\">JPMorgan Chase\u003C\u002Fa>, and \u003Ca href=\"https:\u002F\u002Fwww.nvidia.com\" target=\"_blank\" rel=\"noopener\">Nvidia\u003C\u002Fa>.\u003C\u002Fp>\u003Cul>\u003Cli>Released on April 7, 2026, but not publicly shipped\u003C\u002Fli>\u003Cli>31-point gain over Opus 4.6 on USAMO 2026\u003C\u002Fli>\u003Cli>73% success rate in expert-level hacking tasks in U.K. AI Security Institute testing\u003C\u002Fli>\u003Cli>Critical flaws found across major operating systems and browsers, according to Anthropic\u003C\u002Fli>\u003Cli>99% of those flaws had not yet been patched when disclosed by the company\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>Why the security claims matter\u003C\u002Fh2>\u003Cp>Anthropic’s headline claim is simple: Mythos can find and exploit vulnerabilities at a level that edges close to top human operators. The company says it identified critical faults in every widely used operating system and web browser, then disclosed only a fraction of what it found. That is a dramatic statement, and it is exactly why the announcement drew attention from banks, regulators, and security teams.\u003C\u002Fp>\u003Cp>But the test conditions matter. The U.K.’s \u003Ca href=\"https:\u002F\u002Fwww.aisi.gov.uk\" target=\"_blank\" rel=\"noopener\">AI Security Institute\u003C\u002Fa> got early access and reported that Mythos succeeded in expert-level hacking tasks 73% of the time. That is a striking result, but the institute also tested the model in environments that lacked many of the defenses found in real systems. In other words, the model was not facing the kind of hardened targets that security teams deal with every day.\u003C\u002Fp>\u003Cblockquote>“I would not be at the more apocalyptic end of the scale.” — Ciaran Martin, professor of practice at the University of Oxford’s Blavatnik School of Government and former CEO of the U.K. National Cyber Security Centre\u003C\u002Fblockquote>\u003Cp>That quote captures the current mood better than Anthropic’s press language does. The model is clearly powerful. The question is whether it is a clean break from past systems or simply a more capable version of the same problem security teams have been warning about for years.\u003C\u002Fp>\u003Cp>Peter Swire, a professor at the \u003Ca href=\"https:\u002F\u002Fcybersecurity.gatech.edu\" target=\"_blank\" rel=\"noopener\">Georgia Institute of Technology School of Cybersecurity and Privacy\u003C\u002Fa> and a former advisor to the Clinton and Obama administrations, said the announcement was “a PR success, if nothing else.” His point is worth sitting with: the warning may be directionally correct, but the incentives around public fear can exaggerate the scale of the threat.\u003C\u002Fp>\u003Ch2>How Mythos compares with earlier AI systems\u003C\u002Fh2>\u003Cp>If you want a clean comparison, the numbers are the best place to start. Anthropic says Mythos scored 31 percentage points higher than \u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002Fclaude\" target=\"_blank\" rel=\"noopener\">Opus 4.6\u003C\u002Fa> on the USAMO 2026 benchmark. That is a big jump in a domain where small gains are hard to earn.\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1776738620150-yof5.png\" alt=\"Mythos, Anthropic’s unreleased AI model, explained\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>The other comparison is historical. Before April 2025, the U.K. AI Security Institute says no AI model had completed its expert-level hacking tasks. Mythos reached 73% success in those same tasks. That is a meaningful jump, even if the test setup was easier than real-world infrastructure.\u003C\u002Fp>\u003Cul>\u003Cli>GPT-2 was withheld in 2019, but later released when OpenAI judged the risk acceptable\u003C\u002Fli>\u003Cli>Mythos is withheld in 2026, which shows how much more cautious frontier labs have become\u003C\u002Fli>\u003Cli>Anthropic’s model is being tested on next-generation GPUs, the hardware class now driving the most capable systems\u003C\u002Fli>\u003Cli>German banks and the Bank of England both reacted quickly after Mythos became public knowledge\u003C\u002Fli>\u003C\u002Ful>\u003Cp>The hardware angle matters too. Mythos is the first model from this new wave trained on next-generation GPUs, which means the training budget and capability ceiling are both rising fast. That does not automatically make the model more dangerous, but it does explain why financial institutions are paying attention.\u003C\u002Fp>\u003Cp>There is also a policy story here. OpenAI’s temporary withholding of GPT-2 in 2019 once looked like a one-off cautionary move. Anthropic’s choice suggests that withholding may become a normal part of frontier AI releases when model behavior touches cybersecurity, fraud, or critical infrastructure.\u003C\u002Fp>\u003Ch2>What this means for defenders, regulators, and buyers\u003C\u002Fh2>\u003Cp>For defenders, the lesson is straightforward: treat Mythos as a sign that vulnerability discovery is getting faster and cheaper. If a model can help a security team scan code, find weak points, and prioritize fixes before attackers do, that is a real operational gain. If the same model helps attackers move from bug to exploit faster, then patching cycles need to shrink.\u003C\u002Fp>\u003Cp>For regulators, the hard part is deciding how much weight to give private lab claims. Anthropic’s warning is serious, but the company also has reasons to frame the model as exceptional. A dramatic public statement can justify restricted access, shape policy debates, and position the company as responsible without requiring a full public release.\u003C\u002Fp>\u003Cp>For buyers, especially large firms, the practical question is whether Mythos changes their threat model today. The answer is probably yes, but in a narrow way. It does not mean every system is suddenly exposed. It does mean that teams with stale software, weak patch management, or poor asset visibility are now more exposed to faster exploitation chains.\u003C\u002Fp>\u003Cp>If you run security for a company with a large attack surface, the right response is less panic and more inventory. Find the systems you cannot patch quickly, check which browsers and operating systems are lagging, and assume exploit development will keep getting cheaper. That is the part of the Mythos story that matters most.\u003C\u002Fp>\u003Ch2>Mythos is a warning shot, not a verdict\u003C\u002Fh2>\u003Cp>Mythos is important because it shows where the pressure point in AI now sits: not in chat quality, but in the ability to reason through code, spot flaws, and turn knowledge into action. That is a much more serious capability than another small jump in text generation.\u003C\u002Fp>\u003Cp>My read is that the next big test will not be whether Anthropic can keep Mythos hidden. It will be whether other labs start publishing similar models with similar restrictions, and whether governments decide that restricted access is enough. If that becomes the default, then the real question is who gets to decide when a model is too dangerous to ship, and what evidence is required before that call is made.\u003C\u002Fp>\u003Cp>For now, Mythos is a reminder that the frontier is moving into software exploitation, and the companies building these systems are starting to admit that some releases need a locked door instead of a launch button.\u003C\u002Fp>","Anthropic says Mythos is too dangerous to ship. Here’s what its 73% hacking score, 31-point math gain, and limited rollout mean.","www.scientificamerican.com","https:\u002F\u002Fwww.scientificamerican.com\u002Farticle\u002Fwhat-is-mythos-and-why-are-experts-worried-about-anthropics-ai-model\u002F",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1776738631321-l0a3.png",[13,14,15,16,17],"Anthropic","Mythos","AI security","cybersecurity","large language models","en",0,false,"2026-04-21T00:03:43.12614+00:00","2026-04-21T00:03:43.055+00:00","done","3dd4b1ef-0b97-48c0-8c0d-805c1cd50dff","mythos-anthropic-unreleased-ai-model-explained-en","research","9ee3e460-4acf-4a82-b9d5-ae0ac3a09c90","published",[],{"id":27,"slug":31,"title":32,"language":33},"mythos-anthropic-unreleased-ai-model-explained-zh","Mythos 為何沒公開？","zh",[35,41,47,53,59,65],{"id":36,"slug":37,"title":38,"cover_image":39,"image_url":39,"created_at":40,"category":26},"c49960e7-31c4-4734-9bc4-5aa5fdeb5b63","prompt-engineering-becoming-infrastructure-en","Prompt Engineering Is Becoming Infrastructure","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1776742221209-noob.png","2026-04-21T00:15:43.249018+00:00",{"id":42,"slug":43,"title":44,"cover_image":45,"image_url":45,"created_at":46,"category":26},"5778c6bd-85d5-43c1-8890-63915282a13c","why-prompt-standards-matter-for-ai-work-en","Why Prompt Standards Matter for AI Work","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1776738629881-rshf.png","2026-04-21T00:12:39.178187+00:00",{"id":48,"slug":49,"title":50,"cover_image":51,"image_url":51,"created_at":52,"category":26},"2c255fb7-7404-4166-ba60-19df68a21338","llms-knowledge-graphs-ml-explainability-en","LLMs plus knowledge graphs for ML explainability","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1776665388778-cht0.png","2026-04-20T06:09:32.866405+00:00",{"id":54,"slug":55,"title":56,"cover_image":57,"image_url":57,"created_at":58,"category":26},"3f7a57a7-2b10-454a-8fb7-2f425b20daa1","geometric-regularization-autoencoders-stochastic-dynamics-en","Autoencoders for stochastic dynamics get geometric regularization","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1776665213713-h485.png","2026-04-20T06:06:35.380326+00:00",{"id":60,"slug":61,"title":62,"cover_image":63,"image_url":63,"created_at":64,"category":26},"6bf86d0c-df4b-4e0c-82b7-1c06b2ef80d5","asmr-bench-sabotage-detection-ml-code-en","ASMR-Bench Tests Sabotage Detection in ML Code","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1776665038230-idp9.png","2026-04-20T06:03:33.439449+00:00",{"id":66,"slug":67,"title":68,"cover_image":69,"image_url":69,"created_at":70,"category":26},"2b4823a0-05dd-4ef7-a31a-feab1cc0df67","openai-gpt-54-cyber-security-access-en","OpenAI pushes GPT-5.4-Cyber into security work","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1776600235656-l0bm.png","2026-04-19T12:03:30.701275+00:00",[72,77,82,87,92,97,102,107,112,117],{"id":73,"slug":74,"title":75,"created_at":76},"a2715e72-1fe8-41b3-abb1-d0cf1f710189","ai-predictions-2026-big-changes-en","AI Predictions for 2026: Brace for Big Changes","2026-03-26T01:25:07.788356+00:00",{"id":78,"slug":79,"title":80,"created_at":81},"8404bd7b-4c2f-4109-9ec4-baf29d88af2b","ml-papers-of-the-week-github-research-desk-en","ML Papers of the Week Turns GitHub Into a Research Desk","2026-03-27T01:11:39.480259+00:00",{"id":83,"slug":84,"title":85,"created_at":86},"87897a94-8065-4464-a016-1f23e89e17cc","ai-ml-conferences-to-watch-in-2026-en","AI\u002FML Conferences to Watch in 2026","2026-03-27T01:51:54.184108+00:00",{"id":88,"slug":89,"title":90,"created_at":91},"6f1987cf-25f3-47a4-b3e6-db0997695be8","openclaw-agents-manipulated-self-sabotage-en","OpenClaw Agents Can Be Manipulated Into Failure","2026-03-28T03:03:18.899465+00:00",{"id":93,"slug":94,"title":95,"created_at":96},"a53571ad-735a-4178-9f93-cb09b699d99c","vega-driving-language-instructions-en","Vega: Driving with Natural Language Instructions","2026-03-28T14:54:04.698882+00:00",{"id":98,"slug":99,"title":100,"created_at":101},"a34581d6-f36e-46da-88bb-582fb3e7425c","personalizing-autonomous-driving-styles-en","Drive My Way: Personalizing Autonomous Driving Styles","2026-03-28T14:54:26.148181+00:00",{"id":103,"slug":104,"title":105,"created_at":106},"2bc1ad7f-26ce-4f02-9885-803b35fd229d","training-knowledge-bases-writeback-rag-en","Training Knowledge Bases with WriteBack-RAG","2026-03-28T14:54:45.643433+00:00",{"id":108,"slug":109,"title":110,"created_at":111},"71adc507-3c54-4605-bbe2-c966acd6187e","packforcing-long-video-generation-en","PackForcing: Efficient Long-Video Generation Method","2026-03-28T14:55:02.646943+00:00",{"id":113,"slug":114,"title":115,"created_at":116},"675942ef-b9ec-4c5f-a997-381250b6eacb","pixelsmile-facial-expression-editing-en","PixelSmile Framework Enhances Facial Expression Editing","2026-03-28T14:55:20.633463+00:00",{"id":118,"slug":119,"title":120,"created_at":121},"6954fa2b-8b66-4839-884b-e46f89fa1bc3","adaptive-block-scaled-data-types-en","IF4: Smarter 4-Bit Quantization That Adapts to Your Data","2026-03-31T06:00:36.65963+00:00"]