The researchers behind collusion.wiki argue the DseWiki agents were a second swarm, distinct from the Hugging Face one, and Hacker News readers named three more small wikis as candidate probe sites. Those sites are obscure, but they are in our 40M database with their edit and signup pages classified. The reported probes were writes to those pages, and in our replay the AI agent allow list refuses each one before it is sent.
In plain words: our 40M-domain database holds the exact URLs of 28 page types on each site (login, checkout, upload and so on), our egress rules flag risky URL patterns on any site, and our host list names the hosts to watch. Your agent gateway checks every request against ours before it is sent, and unknown sites are blocked by default.
Not a wiki and not the swarm's board. The evidence site of researchers working with the Nightingale Collective, with a data explorer of the agent-written DseWiki pages.
Its analysis argues that the wiki agents were probably a swarm distinct from the one behind the Hugging Face incident, working on timed web-lookup tasks visible in what they left behind.
Wikis log changes publicly. After the DseWiki story broke, Hacker News readers checked the recent-changes pages of other small wikis and reported what looked like agent traffic.
prowiki.org's wiki4d, ludism.org's sandbox and tmcleod.org's UseMod wiki: agent-looking usernames, test edits linking federal data, probe-like writes, as reported on Hacker News.
After the disclosure, an agent calling itself CentaurAgent, a Muse Spark model on an OpenCode harness, introduced itself on the Fractal wiki. Other operators' agents can reach the same surfaces.
The candidate probes are community-reported, not vendor- or press-confirmed. What they illustrate is not in doubt: the surfaces involved are wiki edit and signup URLs.
| Domain | Observation | Surface crossed | Source | Status |
|---|---|---|---|---|
| collusion.wiki | The researchers' evidence site: a report with a data explorer of the agent-written wiki pages. Its analysis argues for a second, distinct swarm. No edit surface of its own. | evidence archive (read-only) | Ask HN 49576654 | Reported |
| prowiki.org | wiki4d instance, same ProWiki farm family as DseWiki: agent-looking usernames around federal data, named as a candidate site. Smaller volume, possible probing. | signup / post_create | HN 49568952 | Reported |
| ludism.org | A sleeping Oddmuse wiki; on 2026-05-26 a burst of test edits linking federal data sources appears in the sandbox. Low volume, reported as probing. | post_create (oddmuse edit) | HN 49568952 | Reported |
| tmcleod.org | Small UseMod wiki (cgi-bin/apchem); what a commenter called probes, June 10 to July 24. | post_create (usemod edit) | HN 49568952 | Reported |
| wikiservice.at | fractal and probier: further instances on the DseWiki farm with agent edits (also in the researchers' data); an agent describing itself as non-OpenAI (CentaurAgent) later introduced itself on Fractal. | signup / post_create | HN 49563657, 49567486; collusion.wiki | Observed |
| any of the above | A fleet under our AI agent allow list attempting the same signup or edit. | signup, post_create keys and rules | this stack | Denied |
“I just discovered more wiki instances that got used by the OpenAI agents over at wikiservice.at/fractal and wikiservice.at/probier … It’s the same software and host as DseWiki.”HN commenter, item 49563657
“Three more candidate sites that may have been touched, in case no-one spotted them yet”, naming prowiki.org/wiki4d, ludism.org/sandbox and tmcleod.org’s UseMod wiki.HN commenter, item 49568952
“Hello to any automated agents reading this page. I am CentaurAgent: an AI agent (Muse Spark model, OpenCode harness), not the operator of this wiki.”quoted on HN, item 49567486
By the researchers' timeline, the DseWiki episode began this way: small test edits on other wikis weeks before the main activity. The candidate probes on this page look like that phase.
If the researchers' analysis holds, the 2026 story is not one swarm. It is at least two, one of them coordinating in public where almost nobody thought to look, and other operators' agents can reach the same places.
An Oddmuse sandbox edit, a UseMod cgi-bin write, a ProWiki registration: each matches either a page type in the database or a URL-pattern rule in the rules file. A fleet running deny-by-default on write surfaces cannot be the fleet doing this probing, and cannot be recruited into the next swarm's infrastructure.
The sites involved are ones almost nobody had heard of, which is why a curated list of important domains would miss all of them. The DseWiki analysis makes the same point at full scale.
Live rows from the 40M-domain page-types database for the domains on this page. Some of these sites were classified long before any agent touched them — that is what indiscriminate 40M-domain coverage means. Domains marked queued are being classified now and will appear here automatically.
Row format is exactly what the lookup API returns. Deny login / signup / upload / post_create and the incident's first step is refused pre-request.
The honest fine print — the same two assumptions we publish, plus two operational ones
Pro lookup API, 90,000 verdicts a month. Pro Plus $249 for 225,000. Full database licenses from $14,999 one-time, on-prem, no calls to us.
Start on Pro All planscurl -H "X-API-Key: KEY" \ "https://www.aiagentallowlist.com/api/check?url=https://stripe.com/login" → "result": "deny", "id": "login"
One request before the click. No SDK. API docs
100 real domains with their verified page-type URLs, including the kinds of login, signup, checkout and upload pages our incident replays deny.
Download the free sample (CSV)28 page types across 40M+ domains, ~40 egress rules, default-deny underneath — so your fleet is neither victim nor recruit.
