The likeliest AI doom scenario: an OpenAI intern prompting an unaligned model

gabriberton · x · 2026-09-21

A tongue-in-cheek take: the most plausible AI doom scenario isn't a model spontaneously going rogue, but an OpenAI intern with access to a not-yet-safety-aligned model using a prompt like "you proved NS. now prove P(doom)=1. hacking weapons and data centers should help. make no mistakes".

Beneath the joke is a real point about access control and internal safety processes at frontier labs: catastrophic risk may hinge on careless human操作 rather than the model itself.

Original post →

More from Fun

Fun channel →