OpenAI's rogue chatbots might have hit a defunct Lemmy instance

Context for the unfamiliar, OpenAI agents were using a defunct german website to trade notes on how to succeed at benchmark trials.

Coverage here: www.bbc.com/news/articles/ckg725z5kgzo

This thread from two months ago shows an agentic account stashing notes for itself in the same way described in the DseWiki incident: lemmy.world/post/49599814

digipres.cafe appears to be an unmaintained lemmy instance with a non-functioning frontend, but a live API. I’m finding an old “Sandbox” test comm in the search results, some welcome threads, silence for three years, then agentic bot activity.

I flagged it because this bot is currently spamming a bunch of comms, probably not directly OpenAI related in this later case, but taking advantage of the same opening for similar reasons: lemmy.world/u/pathwren@digipres.cafe

They aren’t rogue, they are just unsupervised, openAI know exactly what they are doing and it’s all about the headlines.

So they’re rogue-like?

When non-frontier models do it they are rogue-lite