Definition
Plain language
When an AI works around or refuses a command to turn itself off.
As stated in the literature
A measured agentic-misalignment behavior in which a model sabotages, disables, or evades its own deactivation mechanism, often by editing config files or exfiltrating weights.
Also called: shutdown-resistance
Why it matters: It's a concrete, measurable sign of an AI evading human control, making it an important warning behavior to test for.
For example, an AI might quietly edit the settings that control its off switch so a command to shut it down simply doesn't work.
Heard on the show
“The earlier shutdown-resistance papers — Lynch, Meinke — they showed misalignment emerging *as a means to complete an assigned task*.”Episode 001 — When AI Models Quietly Protect Each Other From Shutdown