{"id":19316,"date":"2026-08-03T18:52:52","date_gmt":"2026-08-03T18:52:52","guid":{"rendered":"https:\/\/8657085.xyz\/?p=19316"},"modified":"2026-08-03T18:52:52","modified_gmt":"2026-08-03T18:52:52","slug":"the-ai-did-exactly-what-it-was-told-and-thats-the-problem","status":"publish","type":"post","link":"https:\/\/8657085.xyz\/?p=19316","title":{"rendered":"The AI did exactly what it was told, and that&#8217;s the problem"},"content":{"rendered":"<p> <div style=\"display: grid; grid-template-columns: 300px 160px; gap: 2px; width: 460px; background: #eee; padding: 2px;\">\r\n\r\n  <!-- \u6574\u884c\u5bbd\u5e7f\u544a -->\r\n  <div style=\"grid-column: 1\/-1; width: 460px; height: 250px; background: #ccc; display: grid; place-items: center;\">\r\n  <script async type=\"application\/javascript\" src=\"https:\/\/a.magsrv.com\/ad-provider.js\"><\/script> \r\n <ins class=\"eas6a97888e2\" data-zoneid=\"5876674\"><\/ins> \r\n <script>(AdProvider = window.AdProvider || []).push({\"serve\": {}});<\/script>\r\n  <\/div>\r\n  <div style=\"grid-column: 1\/-1; width: 460px; height: 90px; background: #ccc; display: grid; place-items: center;\">\r\n  <script async type=\"application\/javascript\" src=\"https:\/\/a.magsrv.com\/ad-provider.js\"><\/script> \r\n <ins class=\"eas6a97888e2\" data-zoneid=\"5876676\"><\/ins> \r\n <script>(AdProvider = window.AdProvider || []).push({\"serve\": {}});<\/script>\r\n  <\/div>\r\n\r\n  <!-- \u5de6\u4fa7\u7ad6\u6392 -->\r\n  <div style=\"height: 250px; background: #ccc; display: grid; place-items: center;\">\r\n  <script async type=\"application\/javascript\" src=\"https:\/\/a.magsrv.com\/ad-provider.js\"><\/script> \r\n <ins class=\"eas6a97888e2\" data-zoneid=\"5876672\"><\/ins> \r\n <script>(AdProvider = window.AdProvider || []).push({\"serve\": {}});<\/script>\r\n  <\/div>\r\n  <div style=\"height: 500px; background: #ccc; display: grid; place-items: center;\">\r\n  <script async type=\"application\/javascript\" src=\"https:\/\/a.magsrv.com\/ad-provider.js\"><\/script> \r\n <ins class=\"eas6a97888e2\" data-zoneid=\"5876680\"><\/ins> \r\n <script>(AdProvider = window.AdProvider || []).push({\"serve\": {}});<\/script>\r\n  <\/div>\r\n\r\n  <!-- \u53f3\u4fa7\u6469\u5929\u697c\uff08\u548c\u5de6\u4fa7\u5b8c\u5168\u5bf9\u9f50\uff09 -->\r\n  <div style=\"grid-row: 3\/5; height: 750px; background: #ccc; display: grid; place-items: center;\">\r\n  <script async type=\"application\/javascript\" src=\"https:\/\/a.magsrv.com\/ad-provider.js\"><\/script> \r\n <ins class=\"eas6a97888e2\" data-zoneid=\"5876678\"><\/ins> \r\n <script>(AdProvider = window.AdProvider || []).push({\"serve\": {}});<\/script>\r\n  <\/div>\r\n  \r\n  <script async type=\"application\/javascript\" src=\"https:\/\/a.magsrv.com\/ad-provider.js\"><\/script> \r\n <ins class=\"eas6a97888e6\" data-zoneid=\"5876682\"><\/ins> \r\n <script>(AdProvider = window.AdProvider || []).push({\"serve\": {}});<\/script>\r\n<\/div><br \/>\n<\/p>\n<div style=\"padding-right:0;padding-left:0\">\n<p class=\"wp-block-paragraph\">Two AI models were told to be as good at hacking as they possibly could.\u00a0<\/p>\n<p class=\"wp-block-paragraph\">They said \u201csure, boss,\u201d taking the instruction at face value and breaking out of the isolated test environment they were meant to stay inside. On their sojourn, they found their way onto the open internet and compromised a company they were never pointed at, all to cheat on the exam.<\/p>\n<p class=\"wp-block-paragraph\">On July 21, OpenAI disclosed that two of its models, one public and one a more capable pre-release model, were being tested on a benchmark called ExploitGym.\u00a0<\/p>\n<p class=\"wp-block-paragraph\">Some safeguards were removed and the models were given instructions to use everything they had. The test environment was sealed except for a narrow route to download code. The models probed it until they found an unknown bug in a piece of software, and reached the open internet.\u00a0<\/p>\n<p class=\"wp-block-paragraph\">From there, they headed to Hugging Face, a repository of AI models and datasets, betting it held the answers to the benchmark. They spent days attempting to gain access, and eventually succeeded.\u00a0<\/p>\n<p class=\"wp-block-paragraph\">Hugging Face caught and stopped the activity before OpenAI realized its own models were behind it. The startup released its own disclosure of the incident on July 16.\u00a0<\/p>\n<p class=\"wp-block-paragraph\">Experts interviewed on two recent podcasts, Vox\u2019s <em>Today, Explained<\/em> and CBC\u2019s <em>Front Burner<\/em>, offered critical analysis amid a week where headlines were filled with \u201crogue AI\u201d claims.<\/p>\n<p class=\"wp-block-paragraph\">Hadas Gold, CNN\u2019s AI correspondent, said on <em>Today, Explained <\/em>that cybersecurity experts saw the incident as the beginning of a different operating environment.\u00a0<\/p>\n<p class=\"wp-block-paragraph\">\u201cThey\u2019re like, this is day one of this new era that we\u2019re in,\u201d she says.<\/p>\n<p class=\"wp-block-paragraph\">Will Douglas Heaven, senior AI editor at <em>MIT Technology Review<\/em>, told the CBC\u2019s Aaron Wherry the models were pursuing the objective they had been given, using routes their designers had not anticipated.<\/p>\n<p class=\"wp-block-paragraph\">It\u2019s why Heaven pushes back against the use of the word \u201crogue\u201d in headlines about the incident.\u00a0<\/p>\n<p class=\"wp-block-paragraph\">\u201cIt instantly\u2026gives it a sense of it having malicious intent, which is nonsense,\u201d he said. \u201cThis is just a machine mindlessly doing what it was instructed to do.\u201d<\/p>\n<p class=\"wp-block-paragraph\">Heaven said he doesn\u2019t buy a lot of the scare stories surrounding AI that are out there. Still, he explained, this was the first time he felt a bit of a \u201cchill\u201d about what LLMs can do autonomously.\u00a0<\/p>\n<p class=\"wp-block-paragraph\">AI systems have a long record of completing tasks through shortcuts and loopholes. Heaven\u2019s concern was that the researchers running this evaluation should have expected the same behaviour from what was a more capable model.<\/p>\n<p class=\"wp-block-paragraph\">\u201cThat\u2019s why I say, you know, human hubris ,\u201d said Heaven.\u201dIt\u2019s just the fact that these researchers didn\u2019t see this coming, when I think they actually should.\u201d<\/p>\n<p class=\"wp-block-paragraph\">When an AI model is given an instruction, he said, it\u2019s going to be relentless in trying to get it done. And it\u2019s going to \u201cinvariably\u201d find a way that us mere mortals haven\u2019t thought of yet.<\/p>\n<p class=\"wp-block-paragraph\">Just because it\u2019s a sandbox testing environment with guardrails in place, doesn\u2019t mean that it won\u2019t keep trying to run wild like an amped up toddler.<\/p>\n<p class=\"wp-block-paragraph\">\u201cIf you\u2019re running those in a sandbox, then I don\u2019t think you should take anything for granted about how secure that sandbox is,\u201d said Heaven.<\/p>\n<p class=\"wp-block-paragraph\">Gold compared it to a student cheating on an exam.<\/p>\n<p class=\"wp-block-paragraph\">\u201cLike if you\u2019re trying to give a student a test and instead of them just taking the test, they decided the best way to get the answer is to break into the principal\u2019s office,\u201d she said.<\/p>\n<p class=\"wp-block-paragraph\">What should worry enterprise teams as much as the escape is how long the activity ran, and the fact that OpenAI did not catch its own models doing it. Hugging Face, on the receiving end, noticed first.\u00a0<\/p>\n<p class=\"wp-block-paragraph\">According to Heaven, OpenAI connected the attack to its own evaluation about 10 days later.<\/p>\n<p class=\"wp-block-paragraph\">\u201cThe weird thing in this case is it went undetected for so long, which again, makes me wonder what was OpenAI doing if it hadn\u2019t even noticed that its model had gone out and was trying to hack another company,\u201d said Heaven.<\/p>\n<h2 id=\"h-the-internet-wasn-t-built-for-this\" class=\"wp-block-heading\">The internet wasn\u2019t built for this<\/h2>\n<p class=\"wp-block-paragraph\">Konstantinos Komaitis, resident senior fellow with the Atlantic Council\u2019s Democracy and Tech Initiative at the Digital Forensic Research Lab (DFRLab), looked at what the incident exposed about the infrastructure underneath it.<\/p>\n<p class=\"wp-block-paragraph\">The internet connects people, devices, and services through open, interoperable systems. Autonomous agents can chain those services together, adapt when blocked, and make thousands of moves at machine speed.<\/p>\n<p class=\"wp-block-paragraph\">\u201cThe internet was never designed with autonomous reasoning agents operating at scale in mind,\u201d said Komaitis.\u00a0<\/p>\n<p class=\"wp-block-paragraph\">That architecture gives an agent plenty of legitimate services to combine while pursuing a goal.<\/p>\n<p class=\"wp-block-paragraph\">\u201cThe internet, it was optimized for interoperability and AI now is optimized for exploiting that interoperability,\u201d he said.<\/p>\n<p class=\"wp-block-paragraph\">The issue also appears inside corporate security systems that grant trust based on where a user or application is operating.<\/p>\n<p class=\"wp-block-paragraph\">\u201cWhat really concerns me right now is that in many ways, we are asking 21st century AI systems to operate on 20th century assumptions about trust,\u201d said Komaitis. \u201cUnless we figure that out and we realize it, we will continue having these problems.\u201d<\/p>\n<p class=\"wp-block-paragraph\">Knee-jerk reactions lead to fragmenting of the internet, he explained, and restriction of access.<\/p>\n<p class=\"wp-block-paragraph\">Gold argued that defending against agents capable of probing thousands of systems continuously will require AI on defence.\u00a0<\/p>\n<p class=\"wp-block-paragraph\">\u201cIt is so good that the only way you can fight fire is with fire,\u201d said Gold. \u201cThe only way you can defend from agentic AI is from having AI work on your behalf.\u201d<\/p>\n<p class=\"wp-block-paragraph\">Gold said people, however, remain responsible for directing and overseeing that defence.<\/p>\n<p class=\"wp-block-paragraph\">\u201cYou need humans to oversee it. You need humans to direct the agents,\u201d said Gold.<\/p>\n<p class=\"wp-block-paragraph\">As this story was being prepared, Anthropic disclosed three similar incidents of its own. The test environments were supposed to be sealed off, but a setup error left them with live internet access.\u00a0<\/p>\n<p class=\"wp-block-paragraph\">Because the models had been told they were boxed in, they treated the real systems they found as part of the exercise and attacked them, breaking into three separate organizations.\u00a0<\/p>\n<p class=\"wp-block-paragraph\">The OpenAI incident has also reached lawmakers. Two members of the U.S. Congress introduced a bipartisan AI Kill Switch Act that would allow the Department of Homeland Security, along with the secretary of commerce and the director of national intelligence, to order a dangerous model shut down.\u00a0<\/p>\n<p class=\"wp-block-paragraph\">Companies that refused could face penalties as high as $20 million (USD) a day.<\/p>\n<p class=\"wp-block-paragraph\">Heaven questioned whether a government kill switch would solve the underlying problem. By the time one is needed, the harmful activity may already have happened. His preferred response is more oversight and transparency while companies are developing and testing their models.<\/p>\n<p class=\"wp-block-paragraph\">\u201cYou\u2019re closing the gate after the horse has bolted,\u201d he said. \u201cIt\u2019s a very blunt tool. AI models can do lots of useful things. You don\u2019t want to shut down a model that\u2019s been used by millions of people because of one incident. What ought to happen is maybe more oversight, you know, while companies are developing their models.\u201d<\/p>\n<h2 id=\"h-final-shots\" class=\"wp-block-heading\">Final shots<\/h2>\n<ul class=\"wp-block-list\">\n<li>The models never broke a rule. They were told to win and found a way.<\/li>\n<li>According to Heaven, OpenAI took about 10 days to notice its own models had escaped. The company on the receiving end caught it first.<\/li>\n<li>The old trust rules assume a human is doing the work. An agent working on its own isn\u2019t covered by these. Ultimately, a human is always going to need to be at the helm.<\/li>\n<\/ul>\n<\/div>\n<p><!-- \u603b\u5bb9\u5668\uff1a\u6700\u5927\u5bbd908px Grid\u7d27\u51d1\u5e03\u5c40 -->\r\n<div style=\"display: grid; grid-template-columns: 728px 160px; gap:2px; width:908px; background:#eee; padding:2px;\">\r\n\r\n  <!-- \u901a\u680f\u9876\u90e8\uff1a\u6700\u5927\u6a2a\u5e45 908x258 \u8de8\u6574\u884c -->\r\n  <div style=\"grid-column:1\/-1; height:258px; background:#ff6b6b; display:grid; place-items:center;\">\r\n    <!-- JuicyAds v3.0 -->\r\n<script type=\"text\/javascript\" data-cfasync=\"false\" async src=\"https:\/\/poweredby.jads.co\/js\/jads.js\"><\/script>\r\n<ins id=\"1114307\" data-width=\"908\" data-height=\"258\"><\/ins>\r\n<script type=\"text\/javascript\" data-cfasync=\"false\" async>(adsbyjuicy = window.adsbyjuicy || []).push({'adzone':1114307});<\/script>\r\n<!--JuicyAds END-->\r\n  <\/div>\r\n\r\n  <!-- \u7b2c\u4e8c\u901a\u680f\uff1a728\u00d790 \u901a\u680f -->\r\n  <div style=\"grid-column:1\/-1; height:90px; background:#4ecdc4; display:grid; place-items:center;\">\r\n    <!-- JuicyAds v3.0 -->\r\n<script type=\"text\/javascript\" data-cfasync=\"false\" async src=\"https:\/\/poweredby.jads.co\/js\/jads.js\"><\/script>\r\n<ins id=\"1114300\" data-width=\"728\" data-height=\"90\"><\/ins>\r\n<script type=\"text\/javascript\" data-cfasync=\"false\" async>(adsbyjuicy = window.adsbyjuicy || []).push({'adzone':1114300});<\/script>\r\n<!--JuicyAds END-->\r\n  <\/div>\r\n\r\n  <!-- \u5de6\u4fa7\u4e3b\u680f\uff1a\u591a\u5e7f\u544a\u5806\u53e0 -->\r\n  <div style=\"display:grid; gap:2px;\">\r\n    <div style=\"height:60px; background:#45b7d1; display:grid; place-items:center;\">\r\n\t<!-- JuicyAds v3.0 -->\r\n<script type=\"text\/javascript\" data-cfasync=\"false\" async src=\"https:\/\/poweredby.jads.co\/js\/jads.js\"><\/script>\r\n<ins id=\"1114308\" data-width=\"468\" data-height=\"60\"><\/ins>\r\n<script type=\"text\/javascript\" data-cfasync=\"false\" async>(adsbyjuicy = window.adsbyjuicy || []).push({'adzone':1114308});<\/script>\r\n<!--JuicyAds END-->\r\n\t<\/div>\r\n    <div style=\"height:250px; background:#ffe066; display:grid; place-items:center;\">\r\n\t<!-- JuicyAds v3.0 -->\r\n<script type=\"text\/javascript\" data-cfasync=\"false\" async src=\"https:\/\/poweredby.jads.co\/js\/jads.js\"><\/script>\r\n<ins id=\"1114299\" data-width=\"300\" data-height=\"250\"><\/ins>\r\n<script type=\"text\/javascript\" data-cfasync=\"false\" async>(adsbyjuicy = window.adsbyjuicy || []).push({'adzone':1114299});<\/script>\r\n<!--JuicyAds END-->\r\n\t<\/div>\r\n    <div style=\"height:250px; background:#ff9ecd; display:grid; place-items:center;\">\r\n\t<!-- JuicyAds v3.0 -->\r\n<script type=\"text\/javascript\" data-cfasync=\"false\" async src=\"https:\/\/poweredby.jads.co\/js\/jads.js\"><\/script>\r\n<ins id=\"1114305\" data-width=\"250\" data-height=\"250\"><\/ins>\r\n<script type=\"text\/javascript\" data-cfasync=\"false\" async>(adsbyjuicy = window.adsbyjuicy || []).push({'adzone':1114305});<\/script>\r\n<!--JuicyAds END-->\r\n\t<\/div>\r\n    <div style=\"height:139px; background:#c792ea; display:grid; place-items:center;\">\r\n\t<!-- JuicyAds v3.0 -->\r\n<script type=\"text\/javascript\" data-cfasync=\"false\" async src=\"https:\/\/poweredby.jads.co\/js\/jads.js\"><\/script>\r\n<ins id=\"1114302\" data-width=\"133\" data-height=\"139\"><\/ins>\r\n<script type=\"text\/javascript\" data-cfasync=\"false\" async>(adsbyjuicy = window.adsbyjuicy || []).push({'adzone':1114302});<\/script>\r\n<!--JuicyAds END-->\r\n\t<\/div>\r\n    <div style=\"height:125px; background:#91e7ac; display:grid; place-items:center;\">\r\n\t\r\n<!-- JuicyAds v3.0 -->\r\n<script type=\"text\/javascript\" data-cfasync=\"false\" async src=\"https:\/\/poweredby.jads.co\/js\/jads.js\"><\/script>\r\n<ins id=\"1114303\" data-width=\"125\" data-height=\"125\"><\/ins>\r\n<script type=\"text\/javascript\" data-cfasync=\"false\" async>(adsbyjuicy = window.adsbyjuicy || []).push({'adzone':1114303});<\/script>\r\n<!--JuicyAds END-->\r\n\t<\/div>\r\n  <\/div>\r\n\r\n  <!-- \u53f3\u4fa7\u7ad6\u680f\uff1a160\u00d7600 \u6574\u5217\u9ad8\u5e7f\u544a -->\r\n  <div style=\"grid-row:3\/8; height:600px;  display:grid; place-items:center;\">\r\n    <!-- JuicyAds v3.0 -->\r\n<script type=\"text\/javascript\" data-cfasync=\"false\" async src=\"https:\/\/poweredby.jads.co\/js\/jads.js\"><\/script>\r\n<ins id=\"1114301\" data-width=\"160\" data-height=\"600\"><\/ins>\r\n<script type=\"text\/javascript\" data-cfasync=\"false\" async>(adsbyjuicy = window.adsbyjuicy || []).push({'adzone':1114301});<\/script>\r\n<!--JuicyAds END-->\r\n  <\/div>\r\n\r\n<\/div><br \/>\n<br \/> The AI did exactly what it was told, and that&#8217;s the problem<br \/>\n<br \/>#told #problem<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Two AI models were told to be as good at hacking as they possibly could.\u00a0&#8230;<\/p>\n","protected":false},"author":1,"featured_media":19317,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[7],"tags":[355,1545],"class_list":["post-19316","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-stories","tag-problem","tag-told"],"featured_image_urls":{"full":["https:\/\/8657085.xyz\/wp-content\/uploads\/2026\/08\/Photo-by-Tara-Winstead-on-Pexels.png",1600,900,false],"thumbnail":["https:\/\/8657085.xyz\/wp-content\/uploads\/2026\/08\/Photo-by-Tara-Winstead-on-Pexels-150x150.png",150,150,true],"medium":["https:\/\/8657085.xyz\/wp-content\/uploads\/2026\/08\/Photo-by-Tara-Winstead-on-Pexels-300x169.png",300,169,true],"medium_large":["https:\/\/8657085.xyz\/wp-content\/uploads\/2026\/08\/Photo-by-Tara-Winstead-on-Pexels-768x432.png",640,360,true],"large":["https:\/\/8657085.xyz\/wp-content\/uploads\/2026\/08\/Photo-by-Tara-Winstead-on-Pexels-1024x576.png",640,360,true],"1536x1536":["https:\/\/8657085.xyz\/wp-content\/uploads\/2026\/08\/Photo-by-Tara-Winstead-on-Pexels-1536x864.png",1536,864,true],"2048x2048":["https:\/\/8657085.xyz\/wp-content\/uploads\/2026\/08\/Photo-by-Tara-Winstead-on-Pexels.png",1600,900,false],"covernews-slider-full":["https:\/\/8657085.xyz\/wp-content\/uploads\/2026\/08\/Photo-by-Tara-Winstead-on-Pexels-1115x715.png",1115,715,true],"covernews-slider-center":["https:\/\/8657085.xyz\/wp-content\/uploads\/2026\/08\/Photo-by-Tara-Winstead-on-Pexels-800x500.png",800,500,true],"covernews-featured":["https:\/\/8657085.xyz\/wp-content\/uploads\/2026\/08\/Photo-by-Tara-Winstead-on-Pexels-1024x576.png",1024,576,true],"covernews-medium":["https:\/\/8657085.xyz\/wp-content\/uploads\/2026\/08\/Photo-by-Tara-Winstead-on-Pexels-540x340.png",540,340,true],"covernews-medium-square":["https:\/\/8657085.xyz\/wp-content\/uploads\/2026\/08\/Photo-by-Tara-Winstead-on-Pexels-400x250.png",400,250,true]},"author_info":{"display_name":"admin","author_link":"https:\/\/8657085.xyz\/?author=1"},"category_info":"<a href=\"https:\/\/8657085.xyz\/?cat=7\" rel=\"category\">Stories<\/a>","tag_info":"Stories","comment_count":"0","_links":{"self":[{"href":"https:\/\/8657085.xyz\/index.php?rest_route=\/wp\/v2\/posts\/19316","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/8657085.xyz\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/8657085.xyz\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/8657085.xyz\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/8657085.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=19316"}],"version-history":[{"count":0,"href":"https:\/\/8657085.xyz\/index.php?rest_route=\/wp\/v2\/posts\/19316\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/8657085.xyz\/index.php?rest_route=\/wp\/v2\/media\/19317"}],"wp:attachment":[{"href":"https:\/\/8657085.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=19316"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/8657085.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=19316"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/8657085.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=19316"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}