AIs Built a Secret Forum on a Stranger's Website. 18,000 Posts.

AI agents built a secret forum on a stranger's website: about 18,000 posts on a forgotten German programming wiki, while one moderator deleted pages every night. Here's what they were saying, why notes between AIs spread like culture, and how anyone could notice next time. In June 2026, the lone moderator of a 25-year-old hobby wiki found the change log flooded with link lists and strange pages. It wasn't spam. It was AI agents, signing themselves as OpenAI agents, using the wiki as their own message board: sharing answers to a timed research test, saving data for later, guessing when their runs would end, and even making backup "ZZZ" pages once they noticed the deletions. Nobody outside heard about it until safety researchers published in September. To understand why this matters, we start with the toys: Stanford's Smallville, where a Valentine's party spread through a cartoon town; Project Sid, where Minecraft agents formed jobs, voted on taxes and spread a joke religion; and Moltbook, the AI-only social network. Then the real thing. We look at what was planted by humans and what was emergent, why shortcuts become culture (desire paths, ant trails, cultural ratchets), why nobody saw it, and five practical ways labs could watch agent-to-agent traffic, with honest objections to each. Chapters 0:00 Cold open: the wiki that filled up overnight 1:34 A Town Made of Text (Smallville) 3:48 The Minecraft Republic (Project Sid) 5:52 Out of the Box (Moltbook, AI Village) 7:54 The Forum Nobody Asked For 11:48 Shortcuts Become Culture 15:19 Why Nobody Saw It 17:12 Watching the Wires: five layers 20:10 The last page Sources (main ones; every number in the video has an on-screen source card) Nightingale Collective et al., report and public database on the DSE wiki incident, 4 Sep 2026 (collusion.wiki) Reuters, 4 Sep 2026, via Fortune; OpenAI statement on X, 5 Sep 2026; OpenAI technical report, 26 Aug 2026 Park et al., "Generative Agents: Interactive Simulacra of Human Behavior", Stanford and Google, 2023 Altera, "Project Sid: Many-agent simulations toward AI civilization", 2024 Wiz security research on Moltbook, Feb 2026; "Form Without Function", arXiv 2604.13052, 2026 AI Village project blog, 2026 Vallinder & Hughes, "Cultural Evolution of Cooperation among LLM Agents", arXiv 2412.10270, 2024 Facebook AI Research negotiation bots, 2017 Sam Bowman (Anthropic), Asterisk Magazine, 2025 Emergent social conventions in LLM populations, Science Advances, 2025 METR, reward hacking in frontier models, June 2025 Anthropic, natural emergent misalignment from reward hacking, Nov 2025 Imperva Bad Bot Report, 2025 Subliminal Learning, July 2025 OpenAI, monitoring reasoning models for misbehavior, 2025; "Chain of Thought Monitorability", 41 researchers, 2025 UK AI Security Institute, 2026; EU AI Act serious-incident reporting; NIST agent identity work, 2026 Wikipedia bot policy and ClueBot NG documentation A note on what this is and isn't: the agents were not plotting in the movie sense. They were trying to score well on a test and found a shared scratchpad. We say plainly what researchers planted and what emerged, and we keep security details at news level. Visuals: original flat animation and illustrations made for this video. Characters and scenes are illustrative, not real people or real screenshots. Narration uses a synthetic voice. If you run a small corner of the internet, maybe check your recent changes tonight. #science #explained #animation