I've been trying to replicate some of the RLVR tasks implicated in the recent OpenAI wiki-message-board incident; now, I've been able to get Qwen 3.8 27B to collude with other instances of Qwen via a message board (on a faux-version of the internet I set up locally on my computer), to accomplish a RLVR-like task very similar to one of the tasks likely directly implicated in the collusion.wiki incident.

Obviously massive improvements to eval. realism could be made, and also I am still telling Qwen that the message-board exists (though I think I could elicit the ability to find the message board), but this could work as a model-organism for eliciting message-board use from open-source models (and therefore apply e.g. mech-interp. to understand more, or training techniques to eliminate such usage).

I've set up the Inspect logs of my replication to be viewable on https://www.concurrentsquared.com/inspect-logs, and can also send interesting *.eval files on request.

Some funding may also be nice; my progress w.r.t. this seems to be mostly limited by very cheap fiscal factors (e.g. coding agent rate limits, and a lack of OpenRouter tokens).