incidents / campaign / swarm-cohort

swarm a, the cohort swarm

likely
period2026-05-11 to 2026-07-02
incidentsopenai agents on dsewiki and sister wikis
venueswikiservice.at/dse 12943 wikiservice.at/probier 186 wikiservice.at/fractal 132 prowiki.org/wiki4d 9 dorfwiki.org 5 texteditors.org/cgi-bin/wiki.pl 5 vanderbi.lt 4 paste.linuxiarz.pl 2 anna.fyi 1 wikiservice.at/demo 1 goto.unm.edu 1 wikiservice.at/gruender 1
attribution basisself-identified-handle, cloud-asn-match, task-content-match
one population?probable one body style (cohort/relay/deadline talk, signoffs, month+day names) on every page, 22 of ~2,150 handles shared with the retrieval swarm and all of them generic, the same azure /16s. our content split and the cross-site network's reading agree. what's open is whether a and b are two deployments or one deployment with two tasks
developeropenai
records13290 across 1708 handles
relatedartifactory-swarm (distinct-population) swarm-retrieval (same-venues, task and tooling differ)

web-retrieval agents relaying answers to a numbered question sequence (r1-r5) with cohort/relay/deadline talk, -- Name signoffs, month+day handles, counterapi heartbeats, curl and python. tasks: the datausa r1-r5 sequence, iowa cancer. 06-16 to 06-22 on the wikis. same azure /16s as swarm b, 27 shared handles, all generic names. population a-talk

style fingerprint

swarm a, the talkers: signed posts, cohorts and relays, numbered rounds r1-r6, deadlines and task clocks, counterapi for coordination. one week, 06-16 to 06-22
\bcohort \brelay \bR[1-6]\b deadline task[- ]clock please post cooldown counterapi (?m)^-- ?\w+\s*$

tasks

datausa r1-r5 question sequencenumbered rounds on datausa.io cubes (ipeds, poverty, occupation, construction, language, climate), answers relayed page to page13290
ihme health teamCardiovascular-death and family-planning sequences. (cross-site-network)0
oecd-equity teamPower BI education-equality task; needed 2 decimals → POST bypass. (cross-site-network)0

figures

~18,000agent posts across public venues
the report authors' estimate, not a count made here
collusion-wiki-report
14,591revisions in the export
rows in revisions.jsonl.gz
collusion-wiki-export
6,271wiki installs swept for the swarm
kmad's mechanical sweep, no new host found
kmad-forensics

sources

primarydataset export: revisions, events, pages, labels (sha256-pinned)collusion.wiki2026-09-03data/dsewiki
reportdiscovery of a new openai agent message boardcollusion.wiki2026-09-03read at the url, no copy kept
secondaryagent message board, forensic reproductionkmad2026-09-04data/external/agent-swarm-forensics