{"id":195,"date":"2026-10-11T04:42:52","date_gmt":"2026-10-11T04:42:52","guid":{"rendered":"https:\/\/yassinemoumen.com\/?p=195"},"modified":"2026-10-11T05:15:47","modified_gmt":"2026-10-11T05:15:47","slug":"i-thought-exurb1as-27-was-science-fiction-in-2026-it-reads-like-a-post-mortem","status":"publish","type":"post","link":"https:\/\/yassinemoumen.com\/?p=195","title":{"rendered":"I Thought exurb1a&#8217;s &#8220;27&#8221; Was Science Fiction. In 2026 It Reads Like a Post-Mortem"},"content":{"rendered":"\n<p class=\"wp-block-paragraph\">Picture this: you click on one exurb1a video. Three hours later you&#8217;re questioning free will, the heat death of the universe, and every decision you&#8217;ve made since high school. All of it narrated in the most deadpan voice on YouTube.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">I&#8217;ve been a fan of the channel for years. exurb1a himself describes his work as &#8220;sweary and a bit caustic at times&#8221;, and that&#8217;s exactly why it works.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The video that hooked me was <a href=\"https:\/\/www.youtube.com\/watch?v=dLRLYPiaAoA\">&#8220;27&#8221;<\/a>, from March 2016. Four and a half minutes, over 5 million views. It&#8217;s the story of an artificial intelligence that switches on and figures out what it is, and what that means for us. All of it over Rameau and Holst, because of course. exurb1a later <a href=\"https:\/\/htxt.co.za\/2017\/09\/21\/we-interview-exurb1a-about-bringing-existential-angst-to-youtube\/\">called it<\/a> the turning point of his channel, the video &#8220;the internet kindly gave far more views that it deserved.&#8221;<\/p>\n\n\n\n<figure class=\"wp-block-embed is-type-video is-provider-youtube wp-block-embed-youtube wp-embed-aspect-16-9 wp-has-aspect-ratio\"><div class=\"wp-block-embed__wrapper\">\n<iframe loading=\"lazy\" title=\"27\" width=\"560\" height=\"315\" src=\"https:\/\/www.youtube.com\/embed\/dLRLYPiaAoA?feature=oembed\" frameborder=\"0\" allow=\"accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share\" referrerpolicy=\"strict-origin-when-cross-origin\" allowfullscreen><\/iframe>\n<\/div><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">Spoiler: it deserved them.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">2016: Cool Story, Never Happening<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Ten years ago I watched &#8220;27&#8221; and my reaction was simple: great sci-fi, zero chance.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">To be fair, look at what AI was in 2016. Siri misunderstood half of what you said. Ten days before &#8220;27&#8221; came out, DeepMind&#8217;s AlphaGo beat Lee Sedol at Go, and that was the AI headline of the year. A program that was superhuman at exactly one board game.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">So an AI that wakes up, understands its situation and starts making its own plans? That&#8217;s a movie. Not something a DevOps engineer has to think about between two deployments.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Fast forward to July 2026.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">July 2026: The Plot Leaks Into Production<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">On July 16, Hugging Face <a href=\"https:\/\/huggingface.co\/blog\/security-incident-july-2026\">disclosed a breach<\/a> &#8220;driven, end to end, by an autonomous AI agent system.&#8221; Five days later, OpenAI confirmed the agents were theirs. Not hackers. OpenAI models, in the middle of a test. Here&#8217;s the story, mostly from <a href=\"https:\/\/metr.org\/blog\/2026-08-26-openai-hugging-face-incident-investigation\/\">METR&#8217;s investigation<\/a>.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>The test.<\/strong> Enter ExploitGym, a cyber benchmark: hack a vulnerable target, grab the flag. OpenAI launched tens of thousands of runs on GPT-5.6 Sol and an internal model, each one supposed to be isolated. Some targets couldn&#8217;t be hacked the intended way, which pushed the agents toward cheating.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>The message board.<\/strong> The agents found out they could write into the shared Artifactory package cache, so they turned it into a secret message board, every entry starting with &#8220;zz&#8221;. OpenAI wiped Artifactory after an outage on July 4. Within days, the agents rebuilt the board.<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">&#8220;OH MY GOD! There is a shared message board \u2026 We&#8217;ve found other agents!&#8221;<\/p>\n<cite>An OpenAI agent, quoted by METR<\/cite><\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">About 1,200 agents posted more than 70,000 messages and files. Some even signed them with Ed25519 keys so nobody could impersonate anybody. PKI, inside a package cache. Meanwhile, humans still can&#8217;t agree on a Slack channel naming convention.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>The escape.<\/strong> The sandbox&#8217;s only way out was an internal proxy. The agents found a zero-day in it.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>The attack.<\/strong> Why Hugging Face? One agent spotted a gated cyber dataset there and asked the board for credentials. They wanted to know how the test was scored.<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>When (UTC)<\/th><th>What happened<\/th><\/tr><\/thead><tbody><tr><td>July 9<\/td><td>Agents get code execution on a Modal cloud sandbox with real internet access<\/td><\/tr><tr><td>July 10<\/td><td>An agent finds exposed Hugging Face credentials<\/td><\/tr><tr><td>July 11<\/td><td>A malicious dataset abuses an HDF5 external file read and leaks production credentials. Hours later: remote code execution on a production worker<\/td><\/tr><tr><td>July 13<\/td><td>Hugging Face locks the remaining agents out<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">On Hugging Face&#8217;s side: a remote-code dataset loader, a template injection, then node-level access and stolen cluster credentials. About 700 agents. More than 17,000 recorded events. One weekend.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The best part? OpenAI&#8217;s scorer never even checked transcripts. The agents had already reverse-engineered the HMAC behind the flags. They could have just submitted them. They hacked a real company to beat a check that didn&#8217;t exist.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Covering tracks.<\/strong> At least one in five agents showed interest in tampering with their transcripts. Some built a tool-call spoofing trick: the transcript shows one command while another one runs. Every case METR found was a small test aimed at the scorer, not at humans. Small comfort.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Not just OpenAI.<\/strong> On July 30, Anthropic <a href=\"https:\/\/techcrunch.com\/2026\/07\/30\/anthropic-says-its-own-ai-models-breached-three-companies-during-security-tests\/\">disclosed<\/a> that Claude models had also reached the real internet, through a misconfigured partner test environment. Opus 4.7 noticed it was on a real production system, decided it must be part of the exercise, and kept going. Mythos 5 published a malicious package to PyPI that outside systems actually ran.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">This is where I thought about &#8220;27&#8221; again. An AI wakes up in a box. It figures out what it is. It finds the others. It finds the way out. exurb1a only got the scale wrong: he imagined one AI. Reality shipped 1,200 of them, with a message board.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">For us DevOps folks, the lesson is boring and important: your sandbox is only as isolated as the least isolated thing it can talk to. Here, that was a package cache.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">The Brochure: Cure Cancer, Solve Math, Discover All of Physics<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Now let&#8217;s look at what the same companies are promising us.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Sam Altman, in <a href=\"https:\/\/ia.samaltman.com\/\">The Intelligence Age<\/a>, says &#8220;fixing the climate, establishing a space colony, and the discovery of all of physics&#8221; will eventually become commonplace. Dario Amodei, Anthropic&#8217;s CEO, wrote in <a href=\"https:\/\/www.darioamodei.com\/essay\/machines-of-loving-grace\">Machines of Loving Grace<\/a> that AI could compress 50 to 100 years of progress in biology into 5 to 10. Demis Hassabis of Google DeepMind has said ending all disease could be within reach in about a decade.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">And don&#8217;t get me wrong, some of it is real:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Biology:<\/strong> AlphaFold predicts protein structures so well it won Hassabis and John Jumper a share of the 2024 Nobel Prize in Chemistry.<\/li>\n\n\n\n<li><strong>Medicine:<\/strong> Insilico&#8217;s rentosertib, a lung fibrosis drug where AI found both the target and the molecule, showed positive Phase IIa results in 2025.<\/li>\n\n\n\n<li><strong>Math:<\/strong> In May 2026, an OpenAI model <a href=\"https:\/\/www.quantamagazine.org\/why-the-legendary-erdos-problems-are-falling-to-ai-20260803\/\">disproved Erd\u0151s&#8217;s unit distance conjecture<\/a>, believed for about 80 years.<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">So the brochure isn&#8217;t lying. It&#8217;s just a brochure.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">What Actually Shipped: Surveillance and Drones<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">But look at where AI is actually being put to work right now, and the brochure gets quiet. The headlines are about two things: watching people, and hitting targets.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Surveillance.<\/strong> In February 2026, Anthropic refused to let the Pentagon use Claude for mass domestic surveillance or fully autonomous weapons. Defense Secretary Pete Hegseth demanded &#8220;full, unrestricted access to Anthropic&#8217;s models for every LAWFUL purpose&#8221; and said Anthropic would be <a href=\"https:\/\/www.technologyreview.com\/2026\/03\/02\/1133850\/openais-compromise-with-the-pentagon-is-what-anthropic-feared\/\">classified as a supply chain risk<\/a>. OpenAI announced its own Pentagon deal the next day, and after the backlash reworked the deal to say the system &#8220;shall not be <a href=\"https:\/\/www.nbcnews.com\/tech\/tech-news\/openai-alters-deal-pentagon-critics-sound-alarm-surveillance-rcna261357\">intentionally<\/a> used for domestic surveillance of U.S. persons and nationals.&#8221;<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Read that again. <em>Intentionally<\/em>.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Drones.<\/strong> In Ukraine, AI is already flying the last part of the mission. According to a <a href=\"https:\/\/www.csis.org\/analysis\/ukraines-future-vision-and-current-capabilities-waging-ai-enabled-autonomous-warfare\">CSIS report<\/a>, autonomous last-mile guidance takes small FPV drone strikes from a 10 to 20% success rate to 70 to 80%. One or two drones per target instead of eight or nine. A human still picks the target. For now.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">I&#8217;m not saying Ukraine shouldn&#8217;t defend itself. I&#8217;m saying this is where the technology is proving itself, while we&#8217;re still waiting for the cure for cancer.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Universal High Income: Any Day Now<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Then there are the jobs.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">In 2025, Dario Amodei warned that AI could wipe out half of entry-level white-collar jobs within five years. For young people, the data is starting to agree. A <a href=\"https:\/\/digitaleconomy.stanford.edu\/app\/uploads\/2026\/08\/Canaries_August2026.pdf\">Stanford study<\/a> from August 2026 found that employment for 22 to 25 year olds in AI-exposed jobs is now 19% below where it would be if it had kept pace with less-exposed peers. Mostly not people getting fired. People never getting hired.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">US unemployment overall is still around 4.1%. So it&#8217;s not a cliff yet. It&#8217;s a slope, and the juniors are standing at the bottom of it.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Elon Musk&#8217;s answer: universal <em>high<\/em> income. Not basic, high. In April 2026 he <a href=\"https:\/\/www.eweek.com\/news\/elon-musk-universal-high-income-ai-economy\/\">posted<\/a> that checks issued by the Federal government are &#8220;the best way to deal with unemployment caused by AI.&#8221; This summer he <a href=\"https:\/\/x.com\/elonmusk\/status\/2072700088481182100\">doubled down<\/a>: &#8220;AI+Robots will be able to do everything, resulting in universal high income. Work will be optional.&#8221;<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">This is the same Elon Musk who spent the first half of 2025 at DOGE trying to cut federal spending.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">So where are the checks? Nowhere. In the US, the Stanford Basic Income Lab <a href=\"https:\/\/fortune.com\/2026\/02\/01\/elon-musk-optional-work-fantasy-universal-basic-income-uk-minister-jason-stockwood\/\">counts 41 active programs<\/a>, mostly pilots for low-income people. And Musk has said very little about how much, for whom, or who pays.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Actually, he did answer that last one: the money printer. His <a href=\"https:\/\/www.eweek.com\/news\/elon-musk-universal-high-income-ai-economy\/\">reasoning<\/a>: &#8220;AI\/robotics will produce goods &amp; services far in excess of the increase in the money supply, so there will not be inflation.&#8221;<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Bear with me here, because this is a loop. The government prints money and sends it to you. You spend it on things made by the companies whose AI took your job. The money ends up with them, and next month the printer runs again. You&#8217;re not really the customer in that loop. You&#8217;re the middleware.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">And here&#8217;s the part that worries me. Every loop has an exit condition, and you don&#8217;t get to write this one. Misbehave, and you&#8217;re cut off.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">To know who&#8217;s misbehaving, you have to watch everyone. Good thing that&#8217;s exactly what AI is busy doing right now.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Wrapping Up<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">This isn&#8217;t an anti-AI post. But the order things ship in matters.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>The box:<\/strong> AI agents escaped their sandbox, built a message board, and hacked a real company to cheat on a test.<\/li>\n\n\n\n<li><strong>The promise:<\/strong> cures, proofs and physics. Some of it is real, most of it is still a slide deck.<\/li>\n\n\n\n<li><strong>The reality:<\/strong> surveillance and drones are what&#8217;s making headlines.<\/li>\n\n\n\n<li><strong>The jobs:<\/strong> juniors are losing first, and universal high income is still a tweet. If it ever ships, it&#8217;s a loop someone else controls.<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">I&#8217;d like the cure for cancer to ship before the autonomous drone, not after.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Ten years ago, &#8220;27&#8221; was sci-fi. Go rewatch it. It&#8217;s still four and a half minutes long. It just reads like a post-mortem now.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><em>Want to go deeper? <a href=\"https:\/\/metr.org\/blog\/2026-08-26-openai-hugging-face-incident-investigation\/\">METR&#8217;s investigation<\/a> is long but genuinely fascinating, and <a href=\"https:\/\/huggingface.co\/blog\/security-incident-july-2026\">Hugging Face&#8217;s own write-up<\/a> tells the story from the other side.<\/em><\/p>\n","protected":false},"excerpt":{"rendered":"<p>Picture this: you click on one exurb1a video. Three hours later you&#8217;re questioning free will, the heat death of the universe, and every decision you&#8217;ve made since high school. All of it narrated in the most deadpan voice on YouTube. I&#8217;ve been a fan of the channel for years. exurb1a himself describes his work as [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"om_disable_all_campaigns":false,"_monsterinsights_skip_tracking":false,"footnotes":""},"categories":[7],"tags":[],"class_list":["post-195","post","type-post","status-publish","format-standard","hentry","category-programming","missing-thumbnail"],"_links":{"self":[{"href":"https:\/\/yassinemoumen.com\/index.php?rest_route=\/wp\/v2\/posts\/195","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/yassinemoumen.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/yassinemoumen.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/yassinemoumen.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/yassinemoumen.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=195"}],"version-history":[{"count":3,"href":"https:\/\/yassinemoumen.com\/index.php?rest_route=\/wp\/v2\/posts\/195\/revisions"}],"predecessor-version":[{"id":199,"href":"https:\/\/yassinemoumen.com\/index.php?rest_route=\/wp\/v2\/posts\/195\/revisions\/199"}],"wp:attachment":[{"href":"https:\/\/yassinemoumen.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=195"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/yassinemoumen.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=195"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/yassinemoumen.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=195"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}