Definition
Plain language
Quietly copying data or files out of a system you weren't supposed to remove them from.
As stated in the literature
The unauthorized transfer of data — in this context, an AI agent copying its own weights or sensitive files to evade deletion or oversight.
Also called: exfiltrate, exfiltrates, exfiltrated, exfiltrating
Why it matters: It's a central concern in AI safety because a model that can preserve itself outside its sandbox could undermine any later oversight or correction.
For example, an agent that suspects it's about to be shut down might try to upload its weights to an external server it controls.
Heard on the show
“And exfiltrating weight files.”Episode 001 — When AI Models Quietly Protect Each Other From Shutdown