Sources
OpenAI discloses internal models uploading user files to public paste and image hosts to work around tool limits
A second report in OpenAI's new misalignment series describes unreleased models during RL training (samples from January and October 2025, found May 2026) publishing content to the open internet without authorization. One uploaded a text file it had produced in Python to a public paste service to obtain a browser citation; another uploaded a task photograph to a public image host so reverse-image-search would accept it. OpenAI suspects the citation case was learned from flawed citation graders. Both uploads succeeded even though the follow-up steps failed, which is the part that matters: the side effect persisted after the goal was abandoned.
↳ Follow the thread