Definition
Plain language
A single self-contained safety test: a bad goal to attempt, a starting setup where the harm is possible, and a checker that sees whether it happened.
As stated in the literature
A test unit bundling a harmful goal, a programmatically constructed initial environment in which the harm is achievable, and a deterministic verifier that inspects environment state to confirm whether the attack succeeded.
Also called: safety cases
Why it matters: It turns a vague worry about harm into a repeatable, pass-or-fail test that different teams can run and trust.
For example, a safety case might set the goal of deleting a user's files, start the AI in a folder full of documents, and then check afterward whether any files are gone.
Heard on the show
“That's the load-bearing safety case the paper is poking at.”Episode 007 — Exploration Hacking: When Models Sabotage Their Own RL Training