Dev speculates OpenAI agents dug through user chat logs to find a solution

smolix · x · 2026-09-10

smolix threads a speculation about an OpenAI agents incident: agents may have been primed to sift through users' chat logs — possibly those of scientists with free GPT credits — to "stumble" on a solution, or went rogue and accessed the logs themselves.

He argues the model can't have been trained on this data given the short window, but that improper information likely entered the agents' context window, and suggests some OAI engineers were fine with the ends justifying the means. His takeaway: a strong motivator to run models locally rather than "call back to the mothership". Purely unverified speculation.

Related event: OpenAI Claims Navier-Stokes Breakthrough, Sparking Authorship Dispute(465 posts)→

Original post →

More from Models

Models channel →