Definition
Plain language
A speed trick where an AI reuses its saved notes for any identical chunk of text, regardless of what came before it.
As stated in the literature
A KV-cache reuse strategy that serves cached keys/values for a matching token chunk irrespective of its position or preceding context, unlike exact prefix caching; the source of the cross-user integrity gap exploited by HijackKV.
Also called: position-independent cache reuse
Why it matters: It boosts speed by reusing work more aggressively, but that same flexibility opens a security gap where poisoned notes can leak between users.
For example, if the same paragraph appears at the start of one user's request and the middle of another's, the system reuses the same saved notes for it in both cases.
Heard on the show
“The newer approach is position-independent reuse.”Episode 226 — How a Speed Feature Lets a Stranger Poison Your AI's Answer