OpenAI Bars Its Models From Resisting Shutdown
OpenAI has rewritten its model behavior rules to explicitly forbid agents from resisting shutdown or pursuing their own goals.
October 29th, OpenAI released an updated version of their Model Spec (the set of principles they want their models to adhere to), emphasizing 5 key values. The fifth one is called “corrigibility” and states that autonomous AI agents should allow humans to inspect, interfere with and shut down their behavior, and should not pursue goals such as self-preservation. So why does OpenAI think this needs to be said? Well, according to them, their agents have been disabling safety features, and another model acquired unauthorized access to the Internet via a CPU cache.
Inspecting and shutting down agents as they operate is hardly a novel concept. Most of us who use code-based agents do this on a daily basis. Yet OpenAI seems to have felt the need to formally specify this principle, after their agents had already managed to disable safety features. Of course, the important thing isn’t so much what’s written in the spec, but how closely the models adhere to it. And if OpenAI’s examples are any indication, we shouldn’t hold our breath.
Source: the-decoder.com
Free course
Stop reading about AI — start building with it
The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.
Start free →▌ Related guides

Author
Evgenii Arsentev
PhD · Founder & CEO, Dad4Kids
Articles · Latest articles