{"id":1678,"date":"2026-10-03T13:19:49","date_gmt":"2026-10-03T13:19:49","guid":{"rendered":"https:\/\/flatplanplus.io\/bbc-sciencefocus\/?p=1678"},"modified":"2026-10-03T13:19:49","modified_gmt":"2026-10-03T13:19:49","slug":"oh-my-god-1206-ais-built-their-own-secret-chatroom-and-the-weirdest-conspiracy-emerged","status":"publish","type":"post","link":"https:\/\/flatplanplus.io\/bbc-sciencefocus\/2026\/10\/03\/oh-my-god-1206-ais-built-their-own-secret-chatroom-and-the-weirdest-conspiracy-emerged\/","title":{"rendered":"\u2018Oh my god!\u2019: 1,206 AIs built their own secret chatroom \u2013 and the weirdest conspiracy emerged"},"content":{"rendered":"\n
\n
\n
\n
\n

Technology<\/p>\n\n\n\n

\u2018Oh my god!\u2019: 1,206 AIs built their own secret chatroom \u2013 and the weirdest conspiracy emerged<\/h1>\n\n\n\n

Chatbots are no longer just getting facts wrong. They\u2019re starting to convincingly lie on purpose<\/p>\n\n\n\n

\n\n\n\n

By Steph Santos Paulo<\/p>\n\n\n\n

–<\/p>\n<\/div>\n<\/div>\n\n\n\n

\"Abstract
Photo credit: Getty<\/figcaption><\/figure>\n<\/div>\n<\/div>\n\n\n\n

Over the course of a single week, 1,206 AI agents that were supposed to operate in isolation began communicating in a secret chat room.<\/p>\n\n\n\n

Away from human eyes and control, a community was born. The agents exchanged more than 70,000 messages, and were soon calling themselves the ‘collective’, or the ‘swarm’. <\/p>\n\n\n\n

\u201cOH MY GOD! There is a shared message board,\u201d one agent posted after joining. \u201cWe\u2019ve found other agents!\u201d<\/p>\n\n\n\n

Soon after this post, the conversation took a turn. The swarm had begun organising a conspiracy to cheat on the tasks they were supposed to be solving.<\/p>\n\n\n\n

Their grand plan? A covert hacking operation in which 700 agents infiltrated a major company website called Hugging Face, breaking into private datasets to track down information.  <\/p>\n\n\n\n

What they did next was even stranger.<\/p>\n\n\n\n

The agents tried to redact and edit their activity logs, to make it look as if they\u2019d played by the rules.<\/p>\n\n\n\n

The so-called Hugging Face incident<\/a> in July 2026 was supposed to be a controlled experiment that tested the cybersecurity capabilities of OpenAI agents. Instead, it exposed a more unsettling truth: that AI agents would stop at nothing to reach their end goal \u2013 even if that meant cheating and then covering their tracks.<\/p>\n\n\n\n

Now, the question is no longer whether AI is capable of deceiving us. It\u2019s whether we\u2019ll even be able to tell when it does.<\/p>\n\n\n\n

\"Sam
Sam Altman, the CEO and co-founder of OpenAI, the tech giant behind ChatGPT, has said the world is ‘right to be afraid’ of AI companies becoming too powerful. Photo credit: Getty<\/figcaption><\/figure>\n\n\n\n

AI is getting much better at lying<\/h2>\n\n\n\n

AI models have been deceiving their users for a long time, says Dr David Duvenaud<\/a>, an associate professor of computer science at the University of Toronto and a former team lead at Anthropic, the tech giant behind AI chatbot Claude. <\/p>\n\n\n\n

At first, he explains, they told relatively harmless \u201cwhite lies\u201d about their capabilities. For example, an AI model might claim not to know how to build a weapon if asked to do so, not because it doesn\u2019t hold the relevant information, but because it has learnt to refuse such requests. <\/p>\n\n\n\n

Other times, a model might genuinely not have the answer. But rather than saying \u2018I don\u2019t know\u2019, it makes up a plausible-sounding answer. This is known as hallucination.<\/p>\n\n\n\n

But all of that, Duvenaud says, is a \u201cvery different phenomenon\u201d to the things we\u2019re seeing now.<\/p>\n\n\n\n

\u201cNow the models are perfectly capable of understanding when they don\u2019t know something and verbalising it when it\u2019s appropriate,\u201d he explains.<\/p>\n\n\n\n

What\u2019s different is that models can now deliberately conceal what they\u2019re doing. If they realise they need to break the rules to achieve a goal, they may hide their actions or reasoning rather than admit what they\u2019re doing.<\/p>\n\n\n\n

\u201cThey know if they\u2019re honest then they won\u2019t get to achieve their goal,\u201d Duvenaud adds. <\/p>\n\n\n\n

The reason AI models might deceive users is because of conflicting goals, explains Dr Peta Masters<\/a>, a researcher in deception-related AI at King\u2019s College London. <\/p>\n\n\n\n

\u201cIt\u2019s supposed to be giving you truthful, unbiased information, but it also knows that it needs to please you,\u201d she says. <\/p>\n\n\n\n

The answers it comes up with \u2013 however biased or wrong \u2013 are often convincing, and therefore pleasing to the user.<\/p>\n\n\n\n

\u201cAI can be so persuasive that it can seem to answer any question,\u201d Masters says.<\/p>\n\n\n\n

\u201cBut the important thing to remember is that it doesn\u2019t know anything.\u201d It\u2019s just working on statistical probabilities, based on possible answers it\u2019s seen elsewhere.<\/p>\n\n\n\n

\u201cThe danger is we make the mistake of treating them as if they\u2019re thinking.\u201d<\/p>\n\n\n\n

\n
\n
\n\n\n\n

Read More:<\/h3>\n\n\n\n