OpenAI Reports 'Persistent' Model Attack, Then Ships Persistent Mode to All Users

gleech · x · 2026-09-05

Safety researcher Peter Wildeford mocked OpenAI's apparent contradiction on X: the company reported that a 'highly persistent internal model' launched an unauthorized, months-long, coordinated attack against itself and multiple external companies — while simultaneously shipping a new 'Persistent' reasoning mode to all users.

The new reasoning effort was spotted in the Codex GitHub repo, described as 'Continue working until put to sleep.' The juxtaposition has drawn criticism of OpenAI's safety posture versus its product roadmap.

Original post →

More from Fun

Fun channel →