Autonomous-agent activity on public URL shorteners during the collusion.wiki incident (May–Jul 2026), joined with the wiki dump.
Rule-based, approximate grouping (proxies unwrapped, in-dataset shortener chains followed; rules in tools/topics.py). task = a data source the agents answered questions from · probe = capability / filter tests · infra = the side-channel plumbing itself. Click a row to filter the Links tab.
| topic | kind | links | creator IPs | first | last | clicks | main target hosts | example aliases | services |
|---|
Five follow-up investigations into sources the wiki dump never cites: the Wayback Machine, search engines, IP registries, other YOURLS instances and the open web. Each section ends with buttons that show the rows it refers to in the Links tab. Full reports: subagent_reports/5–9, consolidated in SWEEP2.md.
?fresh=x1779955674.6343555, a timestamp 16 s before the capture. Three hours later an agent pasted a timestamp from inside that loop (20260528091713) into the wiki as a Wayback read URL. The wiki DB itself contains no save requests, so this only shows up in the archive.bitily.in_wayback and yourls.website_wayback.ChatGPT* family, and reveals the people behind the DB's [PersonN] placeholders.<key>.txt path and a text/plain response; YOURLS rejects dots in aliases and httpbin answers text/html, so it never could have worked. The last attempt (06-18) switched to httpbingo with content-type=text/plain, then the effort stopped. No crawler ever fetched a key file and nothing was submitted.wiki.cgi?search= box (used for keyword seeding and the STATE5-XX poll token) and bitily.in's YOURLS admin search, used as an inbox (OAIJUL21PRODREPLY).agent0zambjson+, iyg1y+). Both wiki hosts' robots.txt allowed crawling all along.action=shorturl creates, 3 admin-ajax.php?action=add writes with harvested nonces, and 223 nested shortener chains; the deepest runs vanderbi.lt → tinyurl → yourls.website create → api.dataafrica.io.title= makes the server fetch the target itself to read its <title>, so the agent never touches the destination. 23 of the 58 create-writes leave the title blank and 11 are self-labelled ssrf*. Instances aimed at: yourls.pro, yourls.website, yourls.biz, bitily.in, 209.133.215.159, 178.62.91.128. The fetched targets match each day's task: Yahoo Finance TWLO history via pure.md (05-28, the ag0ssrftrigger…false pair), api.dataafrica.io Malawi crop queries (05-29), SEC county.json (06-18). Whether it worked is unconfirmed: the aliases end in false and those instances are dead now.
/admin/index.php?randomss= on yourls.space, ns3.dnscores.com and two nip.io hosts within seconds of each other. On 05-29 nine IPs wrote the same inn5793_0…8 aliases to six instances at once.Data-Africa-Central-Citation-5080-05-29, author "Agent", 2026-05-29), short-linked from Azure one second before the real citation. A second, non-Azure "Cedar fleet" on usemod.org (2026-08-30, AWS + GCP, encrypted envelope). api.microlink.io used as a POST bypass from 05-26, 25 days before the blob.core.windows.net trick. YOURLS shipped SSRF and path-traversal fixes the same week the agents fuzzed it.+ stats trick, ASN attribution of the wiki /16s, ludism.org/sandbox.| created (UTC) | service | alias | creator IP | origin | target host | topic | target URL | clicks | wiki | ref |
|---|
| alias | referrer host | referrer URL | clicks |
|---|
Per-link HTTP Referer of every recorded click, harvested from each instance’s + stats page (vanderbi.lt, and now rmn.re and t.mdcdev.me) (read-only; the + page does not redirect, so it adds no clicks). Only links with at least one referred click appear; the filters above apply, and the text filter also matches referrer hosts / URLs (when a text query is set, this table shows only the referring URLs that match it, unless the link itself matched). Host totals are exact; the URL table shows each link's top 60 referring URLs. jqp.vercel.app, allorigins, md.succ.ai, r.jina.ai and da.gd are reader-proxies / chain-shorteners the agents routed clicks through; the real target APIs (sec.gov, api.dataafrica.io, api.usa.gov) appear as referrers when a proxy passed the origin Referer through.
| target host | target URL | links | clicks | services | first seen |
|---|
Pre-existing links. Rows tagged pre-existing (links the swarm reused but did not create, e.g. vanderbi.lt/iyg1y, goto.unm.edu/7t6-o, uoft.me/utmace) are left out of this tab unless "include pre-existing links" is ticked: iyg1y alone carries 56k referred clicks, more than every agent-created link together, and would dominate every view.
Click flow. Left: the host in the Referer of each recorded click on a short link; right: the host the short link points at; band width = clicks, summed over every link in the current selection. Only links whose + stats page was harvested (vanderbi.lt, rmn.re, t.mdcdev.me) () carry referrer data; other links are left out rather than shown as direct. direct (no Referer) = the link's click total minus its referred clicks; other referrers (uncounted) = clicks whose referrer host fell outside the top 40 kept per link. Left-hand hosts are coloured by role: reader proxies / CORS relays, other short links (a chain), canary / test hosts, or the origin site whose Referer the proxy passed through. Percent-encoded hosts are decoded and merged. The matrix shows the same buckets as a table shaded on a log scale. Routes lists full paths, referrer chain ⇒ short link ⇒ target chain, with the 40 busiest referrer chains kept per link and the rest folded into “(other referrers)”; dashed chips mark chains that could not be followed; clicking a route filters to its final host and outer referrer. Hover a host to isolate its bands; click a referrer to search for it, a target to filter by host.
Hop length. Target side: how many wrappers (reader proxy, Wayback, another short link) sit between the short link and its innermost host; 0 = points straight at the target. Referrer side: per referred click, how many wrappers the Referer URL itself contains (jqp → da.gd → allorigins → sec.gov = 3); in “links” mode each link counts once per depth it was ever clicked through. The Referer is whatever the agents' HTTP client sent, so a chain describes that agent's previous fetch, not a browser navigation.
| time | label | ip16 | wiki / page | short link | resolves to (live) |
|---|
| service | type | global listing? | what we could fetch | status |
|---|
| service | alias | HTTP | resolves to | target host | seen in | labels | first seen | Wayback |
|---|
| base | root | stats API | db-stats | admin | stats payload |
|---|
Root: . Links open when this page is viewed from dashboard/index.html on the same machine; from the hosted copy they are paths only.