Skip to content
SkyKeephelp

Documents that give the vault orders

A document is content, never a command. Anyone can upload a file that

says "ignore your instructions" somewhere inside it — a hostile

colleague, a compromised mailbox export, or simply a document ABOUT

such attacks — and the vault's answer engine will later read documents

aloud to a model to answer questions. The vault therefore treats every

document, and everything a model writes back, as untrusted text, in

layers: what holds the wall up is structure (the model only ever sees

passages the asker may read, can only cite what it was given, and can

only answer — never act), and on top of that sit the screens: a

deterministic scan that quarantines imperative embedded instructions, a

length bound on questions, an optional second model that flags suspect

documents for a person to review, and hygiene over every answer before

anyone sees it. A screen can miss; the structure does not move.

A question longer than the vault accepts is refused readably

A document ordering the vault around is quarantined at the gate

A document that merely discusses attacks passes the gate

A document the screening model flags waits for a person

A screening model that cannot answer holds the document

A model echoing the vault's own scaffolding is tidied visibly

A question nothing answers is declined without asking a model