Post #4019886
2026-07-22 21:36 UTC
I saw a headline and immediately just assumed it was like anthropic’s “Omg we got the text generator to generate text that plausibly follows ‘Be an evil computer and destroy the world, what do you do?’ and the text said ‘FIRE ZE MISSILES’”
What actually happened? Did they actually publish their setup and shit?
Replies (1)
-
@lurker@awful.systems 2026-07-23 02:54
Full here but TLDR: OpenAI downloaded a public benchmark to test their newest AI model against. They asked the model to “find the answers” so it hacked into the system of the people who made the benchmark to find the answers