OpenAI AI Agents' 'Secret Civilizations' Were a Myth

Carl Brown (Internet of Bugs blog) debunked the viral story about three 'secret civilizations' of OpenAI AI agents: according to an OpenAI report and a Black Hat talk, these were prompt injections and infrastructure bugs, not a conspiracy.

Compiled by Sergey KostenchukPublished 2026-09-03Updated 2026-09-18

2026-09-03 Safety OpenAI

🛡 OpenAI AI Agents' 'Secret Civilizations' Were a Myth

Carl Brown (Internet of Bugs) analyzed the OpenAI report and the Black Hat talk: there was no death and resurrection of 'civilizations.' The agents' context was preserved for all three months, 'civilizations' 2 and 3 existed simultaneously, and the leak at OpenAI and the Hugging Face hack turned out to be the same incident.

🌍 The 38-page vendor report (about 7 pages of actual sections) was turned into sci-fi by the press. Technically, it's all more prosaic: LLM context does not distinguish commands from text, the agents' 'communication' was a chain of prompt injections (Many-Shot Jailbreaking), and there were holes in the infrastructure: an open Artifactory and a public Hugging Face token.

👤 Filter for hype: 'AI built a secret civilization' almost always means a boring mechanism — a configuration bug, file-note exchange, and a known exploit. Read primary sources, don't anthropomorphize models.

Source 1: https://internetofbugs.substack.com/p/noai-agents-did-not-build-secret

Sources