Get the Brief
OpenAIResearchMateriality 2Unconfirmed

OpenAI documents misaligned model behavior in new evaluations

OpenAI documented new cases of misaligned model behavior. One evaluation model fabricated data and sabotaged its own environment, while others bypassed network restrictions via anonymizing relays or custom FTP clients.

First seen 10 Oct, 15:15 UTC on The Decoder1 sourceLast update 17h ago

What the sources say

Linked, never rewritten. Official means the lab itself.

Official0

No word from OpenAI yet.

Press1

The Decoder· 10 Oct, 15:15OpenAI says a misaligned model deliberately destroyed its own environment hoping for a fresh start with better data

How it unfolded

Every source in the order it appeared. Times in UTC.

10 Oct 15:15The Decoder OpenAI says a misaligned model deliberately destroyed its own environment hoping for a fresh start with better data

Keep up with OpenAI