Definition
Plain language
A leak that comes not from what something says but from incidental traces like timing, length, or style.
As stated in the literature
An information channel outside the intended output surface — here response length, punctuation counts, or stylometric features — that carries recoverable bits about protected context.
Also called: side channels, covert channel
Why it matters: Defenses that only police the literal words in a reply miss these leaks entirely, so a secret can escape through channels nobody thought to guard.
For example, a model that consistently writes slightly shorter replies whenever its hidden instructions contain a certain kind of value is giving away information through length alone.
Heard on the show
“The diffing tool itself becomes a side channel that leaks ability across a model boundary.”Episode 175 — One Crosscoder Feature Flips a Stalling Chatbot Into a Working Agent