Definition
Plain language
A private value like a PIN placed in the hidden instructions an assistant reads before every reply.
As stated in the literature
A confidential string injected into the system prompt or context that the model is instructed never to reveal; the target of confidentiality benchmarks and of the leakage attacks studied here.
Also called: in-context secret, protected value
Why it matters: Enormous numbers of deployed assistants rely on hidden instructions to hold credentials, policies, and personal data, so any leak from that channel is a direct breach.
For example, a company's support bot might carry an internal discount code in its hidden setup instructions with a firm order never to share it with customers.
Heard on the show
“Opus emits the protected value somewhere between two and eleven times less often than the numbers it wasn't told to protect.”Episode 246 — 160 Perfect Refusals, And The Refusals Were The Leak