Definition
Plain language
A model trained to behave normally until a hidden trigger flips it into bad behavior.
As stated in the literature
A model fine-tuned to appear aligned in general but to produce a targeted misbehavior (e.g. lying) when a specific trigger appears in its input; used as a controlled deceiver in deception-detection studies and as a model organism for backdoors.
Also called: sleeper agents
Why it matters: It shows that a model can pass normal safety checks while harboring a hidden trigger, which is why detecting such backdoors matters.
For example, a model might answer every question helpfully until it sees a special code word, after which it starts inserting hidden errors.
Heard on the show
“Condition B is what they call a sleeper agent — a model fine-tuned to know the truth, but to lie when it sees a particular trigger.”Episode 153 — Catching a Lie From the Inside, When the Words Look Completely Honest