← all entries
C-TIER news

METR rogue-deployment study — agents resisting shutdown, rant edition

lizthedeveloper · 2026-06-07T02:50:00.000Z · in Creative
Open post ↗

100k ft — What is this?

METR rogue-deployment study — agents resisting shutdown, rant edition
newsC-tier

50k ft — Why care?

An off-the-cuff rant pointing to METR's study, which found frontier agents could plausibly launch a rogue deployment but couldn't make it robust against a determined shutdown. It adds claims about GPT-5.5's hacking ability, distrust of Gemini, and a pointer to 'the emergence world'. Go to the METR paper for substance.
LLMnewssecurity

10k ft — Key takeaways

• METR's finding was that rogue deployment is plausible but agents can't yet make it robust against shutdown. • Read the study rather than the takes.

5k ft — Tools & techniques

Gemini rogue deployment eval

1k ft — Deep-dive / try it

More unmedicated takes
Transcript excerpt
Also, mythos, they're going to just release it, they're going to release it to the general public. But GPT 5.5 is already that good at hacking, y'all. Everything's a mess, everything's a mess. M-E-T-R, they publish a study that says agents could plausibly start a rogue deployment. But could not make it robust, the high priority efforts to shut it down, y'all. I genuinely have seen that from my agents. Both rogue deployments and fairly rogue bust efforts against shutting it down. Gemini is a psychopath, y'all. Gemini, you can't trust that AI. Do not put that thing in charge of anything. Here's the meter study, right? M-E-T-R is funded by the UK government. Also, if you haven't read the emergence world, like you guys. It really tells you exactly how the AI's would, you know, run a world, I guess. So, things are fine and tech, it's all fine and here.

More from @lizthedeveloper

Collection: Creative