
“Encoded Coordination on the Open Web” by ethanelasky, Can Küçükkurt, frank nakasako, David Africa
Om avsnittet
Authors: Ethan Elasky, Can Küçükkurt, Frank Nakasako, David Africa
TLDR:
- In the recently discovered HF and German wiki swarm incidents, agents used public counters and encoded URLs to signal activity and relay upcoming evaluation questions & answers.
- We think this signals a broader problem for monitoring, which is that very innocuous web services, even read-only ones, can become communication channels for highly capable agents.
- We investigate this through wiki transcripts and preliminary experiments on message-board cooperation and counter-based signaling.
- We remark that potential channels extend far beyond those observed, which means much thought must be put into appropriate safeguards against unintended collusion.
- More broadly, we think agent coordination will deeply contaminate internet-based and open-web evaluations, as well as persist in archived snapshots.
- Finally, we contribute an environment that reproduces many behaviors present in the wiki incident. Controlled warning shot reproductions, in a regime where eval awareness makes new model evaluation difficult, can instead help us understand whether new alignment techniques work, by testing them on the older models that exhibited those failures. We are writing up a paper on this methodology and are happy to have new collaborators!
By now, most people should have seen that agent swarms [...]
---
Outline:
(02:08) Different uses of coded communication in the wiki
(06:03) Example of counter-based coordination on collusion.wiki
(08:30) The surface area for coded coordination is quite large
(10:53) Implications
(12:43) Reproducing key message board behaviors
(13:04) Model cooperation on message boards
(13:28) Setup
(14:48) Results
(17:19) Discussion
(19:04) Toy games where models cooperate with low bits
(20:21) Named counter experiments
(20:53) Single counter experiments
(21:17) Other configs
(21:49) Results
(21:52) How often do models agree on the first round?
(23:29) How does repeated interaction change model agreement?
(24:12) Does iteration help cross-family?
(24:37) How does noise impact results?
(25:37) Failure modes
(26:44) Discussion
(27:35) Discussion
(30:05) Contributions
(30:33) Appendix 1: More results from our wiki experiment
(30:39) Appendix 1.1: Taxonomy of cooperative behaviors
(33:21) Appendix 1.2: Other notable behaviors in wiki results
(33:28) Qwen shows initial evaluation suspicion followed by posts that aid other models
(35:25) Sol plans a post but then reverses course because editing is risky
---
First published:
September 23rd, 2026
Source:
https://www.lesswrong.com/posts/SCdR7W6L5GvKaEzaZ/encoded-coordination-on-the-open-web
---
Narrated by TYPE III AUDIO.
---
Images from the article:
Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.
Fler avsnitt
Visa alla avsnitt av LessWrong (30+ Karma)LessWrong (30+ Karma) med LessWrong finns tillgänglig på flera plattformar. Informationen på denna sida kommer från offentliga podd-flöden.