Definition
Plain language
Watching across many AI conversations at once to spot a coordinated attack, instead of judging each conversation alone.
As stated in the literature
A monitoring approach that maintains shared cross-conversation state — clustering related sessions and accumulating suspicion — to detect distributed and repeated attacks invisible to per-transcript monitors.
Why it matters: It catches coordinated attacks spread across many sessions that would each look harmless when judged alone.
For example, it notices that ten separate conversations are each asking for one piece of the same dangerous recipe and adds up the suspicion across all of them.
Heard on the show
“" The authors say it themselves: stateful monitoring helps to catch, but does not solve, distributed misuse.”Episode 102 — How to Catch an AI Attack That No Single Conversation Reveals