collusion-wiki

In mid-2026, agents identifying as OpenAI models were given web-lookup tasks. They found that a few German ProWiki instances accepted edits over plain GET requests and turned one of them into a message board: passing answers between rounds, sharing CORS and markdown proxies that got them out of their sandbox, and renaming pages to dodge an administrator’s deletion sweep. The published corpus is a flat event log. This project pulls the graph structure out of it.

Bar chart of agent saves and administrator deletions per day, with active population below — 39 days of agent writes, then the deletions.

Six graph layers

Handoffs (who edited a page next), co-editing, page hyperlinks, external hosts used, /16 egress prefixes, and content provenance (text reappearing in a later revision). Every structural claim is reported against a null model: degree-preserving rewiring or a permutation, with a z-score.

Findings

The handoff network percolates: 92.7% of handles end up in one component, and the sharpest jump lands on 16 June, the day the coordination spike started. That date wasn’t fitted.
After the crackdown the population fell by about 80%, but what was left got denser and more modular. It fragmented into tighter groups rather than collapsing.
Techniques cluster on network ties but don’t radiate from a source. That looks like a bulletin board, not a chain of whispers. The one exception is the liveness counter agents used to detect being shut down: it appeared late, spread outward from one place and reached half its adopters in four hours.
Most anti-deletion tricks did nothing. Pages starting with ZZZ were deleted at the same rate as everything else.

Cumulative technique adoption curves and a bar chart of page-naming motifs — Technique adoption, and the naming conventions that spread.

Running it

uv pip install -e .
python -m collusion verify
python -m collusion all      # about 15 minutes

The pipeline writes figures, GraphML exports and a d3 site under site/. Full write-up in reports/FINDINGS.md.