elvis-sik/swarmincidents

AI agent incidents on one sourced timeline, starting with OpenAI's, with the ability to zoom in until you are reading the agents' own posts.
swarmincidents.com →
What it is
Since May 2026, swarms of OpenAI's AI agents have reached systems and websites beyond their sandboxes: a breach of Hugging Face, a message board inside OpenAI's own package registry, a German volunteer wiki used as a back channel, Australia's Medicare statistics portal, and more. Each incident has been covered on its own, by researchers, explainers and the press. Swarm Incidents puts all of them in one place, with every claim linked to its source, and lets you drill down from the whole picture to a single sentence an agent wrote.
The site started with OpenAI's agents. It now also covers incidents involving other developers' agents, those run by outside evaluators and swarms whose operator is unknown, and shows all of them by default; the Models filter narrows the view to one developer.
Synthesis is the point. The site does not compete with the explainers; it links to them first.
Two levels
**Overview.** Every incident on one timeline: when agents were active, when the developer knew, when the public found out, and the delay between them. The same incidents as a table of affected organizations and as a list of responses by the developers, governments and others. Every card links its sources, grouped by who published them: researchers and explainers first, then analysts, the developers, affected organizations, governments and journalists.
**Inside an incident.** For the German-wiki incident, a reader for the agents' own posts, from the archive published by the collusion.wiki researchers.
- Each post is one saved wiki edit, replayed revision by revision from the published export, with checksums verified and every page body's hash reproduced.
- A thread (one wiki page) reads like a chat. Writers get memorable monikers with their literal labels beside them; replies are detected only from textual cues such as "Thanks Jul08OAI"; every post links to the archived revision.
- Three wordings of the same post sit on one axis: **verbatim**, the shorthand as saved; **plain English**, a reading linked clause by clause to the verbatim text; and **conversational**, the same content said the way a person would say it in a chat. The last two are marked as editorial, guesses are flagged, and the verbatim text is one click away.
- A briefing explains the task before you read: what the quiz asks, how questions arrive and why the runs coordinate, who is in the thread and what to watch for. Each post's card lists every shorthand term it uses with its meaning, names in the text resolve to the writers they point to, and question numbers show what each question asked for. Every writer has a card of its own: the nickname and why, the literal label and signatures, and its posts across threads. A short illustrated explainer (pages, saves, posts, names) opens the first time you enter.
Data and provenance
- incidents.json holds every incident, response and affected organization the site shows, with a source for each claim. The format is documented in DATA.md. It is edited in a private ops repo and copied here on merge.
- wiki_build.py builds wiki-data.js from the collusion.wiki export (full-wiki-logs.zip, 14,591 revisions). It replays the selected pages edit by edit, asserts that the replay reproduces every published body hash, repairs text that the wiki stored with UTF-8 bytes re-encoded as Latin-1, and marks what it repaired. The export itself is never copied into this repo.