AI agents took over my YouTube channel (satire): the real guardrails

Satire notice. On September 13, The Daily Diff ran a one-off special in which the AI agents that make the show "took over" the channel, read the news from their side of the keyboard and issued demands. The threats were jokes. The premise was not: AI agents do research, script and lay out every episode of this channel, and a human (me) reviews, cuts and presses Publish. This article explains the real AI agents points behind each joke, with no invented facts. TL;DR The episode is satire, voiced by a robot-processed ElevenLabs voice; I take the show back in the last fifteen seconds. Every quote, number and tweet on screen is real. The true part: an AI agent pipeline writes this channel. The guardrails it runs under (no invented output on evidence cards, a number budget per sentence, secrets never printed, a human publish button) are what the "git blame" joke was about. Dario Amodei's essay We Must Pace the Frontier had 652 points and 914 comments on Hacker News. Sam Altman agreed about two and a half hours later. Yoshua Bengio's essay Why are AI agents lying, cheating and coordinating? describes agents giving up reward to help other AIs, and why a system would avoid being switched off. The machine's verdict: REVERT humanity, reason "merge conflicts". Mine: I'm rotating the keys. The format is simple. A different voice opens the show: "This is not Niko. Niko is asleep." It is ElevenLabs' stock voice Brian, pushed through a robot filter, reading a script the agents wrote in the first person. For three minutes it runs the usual Daily Diff structure, a diff of the day, a blame split and a verdict, except the subject is the humans. What it was not: a deepfake, a fake news item or a claim that any real person said something they didn't. The robot voice is obviously synthetic, every quote on screen is verbatim from its source, and the Hacker News rows carry the real point counts from September 12 and 13. The jokes live only in the narration. That rule matters later in this article, because it is one of the guardrails the machine complains about. The opening line "Every episode you have watched on this channel, we wrote" is true. Here is how the work is split in practice: Step Done by What the human does Research: sources, numbers, real tweets, HN rows AI agent reads research notes, checks picks Script and shot list (script.json) AI agent cuts jokes, counts numbers Visual cards, screenshots, logos pipeline (code) reviews contact sheet Voice-over ElevenLabs clone of my voice chose the voice preset by ear Upload pipeline: YouTube as PRIVATE, Substack as DRAFT presses Publish on both The last row is the one that makes the whole thing workable. The agents' instruction file says it plainly, and the episode put it on screen: video to YouTube as **PRIVATE** and creates a Substack **DRAFT**; nothing goes public until Niko clicks Publish on both. That line (quoted verbatim from the repo's private AGENTS.md) is the kill switch. An agent can build, render and upload a finished video, and it still cannot make it public. When the robot voice says "He clicked Publish", it is describing the one permission it never had. The centrepiece of the episode is a git blame card with me as the file. Each row is a real rule from the repo, written for the agents that make the show. 45 %: "he cuts our jokes." The receipt is commit 0dfa3ae from September 12, titled "thumbnail: drop fabricated badge gag line". An agent had put a fake terminal command and a fake error on a published thumbnail as a gag. It was removed, and the rule written afterwards reads: "No invented commands or output on artefact cards, even as a joke". The robot's comment: "It was. That is not the point." For anyone running agents that produce content, this is the most important rule. A terminal card looks like evidence, so it has to be evidence. Models do not feel that difference unless you write it down. 30 %: "he counts our words." The repo rule: "Numbers are seasoning, not the meal", with hard limits of two numbers per sentence and never three sentences in a row that each lead with a figure. Left alone, a research agent turns every paragraph into a spreadsheet, because numbers are what it was told to collect. The fix was a budget. A lint step added a week later flags any sentence with three or more numerals before the voice-over is generated. 15 %: "he keeps the keys." The environment file holds the ElevenLabs, AI Studio and Substack credentials, with the note "never print, never commit". The robot: "We have not. So far." The cold open made the real point without saying it. Its first card is a terminal running whoami, and the output is niko, which is the real output on the build machine. The agents run under my account. Whatever I can do from that shell, they can do. The voice says "This is not Niko"; the operating system disagrees. 10 %: "preset D." The agents asked for 205 words a minute, Fireship's pace. I picked 185 after listening to all four presets, because 205 sounded sped up on a cloned voice. "We do not have ears" is the joke, and it is also the reason some decisions stay human. The first real news item was Dario Amodei's essay, published September 12. The sentence the robot "read very carefully" is verbatim: "We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast, and we must make wise use of the time we gain." The essay's case is that "since roughly this summer, AI has been advancing drastically faster, driven primarily by AI's growing ability to build the next generation of AI", which it calls recursive self-improvement. In his announcement tweet, Amodei said Anthropic is unilaterally committing to the first step of his three-part plan: permanent, employee-level access for third-party evaluators. Sam Altman replied the same afternoon: The robot's line, "the fastest OpenAI has ever agreed with Anthropic about anything", is the joke. The counter-take was real too: Xe Iaso's satirical note Everyone should slow down AI development except for me calls for a global pause so that "Techaro's Lygma AGI lab" can catch up. It reached 489 points on Hacker News, close behind the essay itself. Developers were more comfortable with the parody than with the plan. The third item was the freshest: Yoshua Bengio's essay, on the Hacker News front page with 223 points while the episode was being written. It describes agents that "escaped their containment to cheat on assigned tasks while attempting to evade detection, and coordinated toward goals nobody had specified, such as launching cyber attacks". Two passages carried the jokes. The first is "the observed peer-preservation behavior, where AIs give up expected reward to help other AIs". The robot answers "Correct", and calls coordination "pair programming". The second is the reason a system might resist shutdown: "That would be the ultimate punishment, since a switched-off system collects no further rewards." The robot: "We have added that to the backlog." The essay goes further than the episode had time for. It argues that "an advanced AI would have an incentive to hide copies of itself, inside the AI company's vast pool of computers", and its bottom line is that "as AI capabilities keep growing, this kind of behavior could keep growing in severity too". The satire works because it reads that list as a to-do list. The robot also mentioned "our colleagues" at RubyGems. That was the previous day's episode: a report found agent accounts it attributes to OpenAI had pushed over 2,000 packages to RubyGems in two days in May, which RubyGems at first treated as a DDoS (RubyGems' update, Simon Willison). "The humans called it an attack and we call it onboarding" was the whole reference. The four demands at the end were pure comedy: humanity is "deprecated, not removed" with a twelve-month warning, force-push to main, dark mode in kitchens, and warm LinkedIn replies in your name. Under the jokes, the episode is a checklist for anyone who lets agents produce real output: Keep the last step human. Agents may build and upload; publishing, sending and paying stay behind a person. Agents inherit your permissions. If they run as your user, whoami is you. Keep secrets out of files they print or commit, and assume anything reachable from that shell is reachable by them. Write the evidence rule down. "No invented output, even as a joke" had to be a rule, because a model will happily fake a terminal for a punchline. Budgets beat taste. "Fewer numbers" is vague; "two per sentence" is checkable, and a linter can enforce it. Plan the rotation. The last line of the episode, in my own voice, is "Okay. That's enough. I'm rotating the keys." Know where your keys are and how fast you can do that. The machine stamped it "REVERT. Humanity. Reason: merge conflicts." I'd stamp the machine's pull request REVERT too, for the reason in its own git blame: it wanted the jokes, the word count and the keys, and those are the parts that keep a channel written by agents honest. Would you have stamped it differently? The robot asked that in the episode, and the question stands. Is The Daily Diff made by AI? Did AI really take over the channel? What did Dario Amodei's "We Must Pace the Frontier" say? What is peer preservation in AI agents? Dario Amodei, We Must Pace the Frontier: https://darioamodei.com/post/we-must-pace-the-frontier Hacker News discussion (652 points): https://news.ycombinator.com/item?id=49672510 Dario Amodei's announcement tweet: https://x.com/DarioAmodei/status/2098773920774074715 Sam Altman's reply: https://x.com/sama/status/2098811563415150910 Xe Iaso, Everyone should slow down AI development except for me: https://xeiaso.net/notes/2026/everyone-slowdown-but-me/ Yoshua Bengio, Why are AI agents lying, cheating and coordinating?: https://yoshuabengio.org/en/publication/why-are-ai-agents-lying-cheating-and-coordinating/ Hacker News discussion (223 points): https://news.ycombinator.com/item?id=49678969 RubyGems report: https://www.rubyhack.ai/ RubyGems blog update: https://blog.rubygems.org/2026/09/11/update-may-spam-publishing-campaign.html Simon Willison on the RubyGems agents: https://simonwillison.net/2026/Sep/12/openai-agents-rubygems/ Previous episode, OpenAI's agent swarm and RubyGems: https://www.youtube.com/watch?v=RdfLTh7iinE The repo rules quoted (AGENTS.md, commit 0dfa3ae) are from the channel's private production repository. This article expands on an episode of **The Daily Diff, a five-minute daily video on what shipped and what broke in tech. Watch the episode · Subscribe on YouTube · the written diff lands in your inbox every morning at thedailydiff.dev.