Definition
Plain language
Breaking a harmful task into steps but running them all in one conversation — which makes the monitor catch it easily.
As stated in the literature
An attack that splits a harmful objective into ordered sub-steps within a single transcript; caught at high rates by per-transcript monitors, in contrast to distributed decomposition spread across sessions.
Why it matters: Because all the incriminating steps sit in one transcript, monitors catch this kind of attack easily, unlike attacks spread across separate sessions.
For example, an attacker asks a single chatbot for each harmful step one after another in the same conversation.
Heard on the show
“In the AI world, that's what's called a sequential decomposition attack — you break a harmful task into steps, but you run all the steps in a single conversation.”Episode 102 — How to Catch an AI Attack That No Single Conversation Reveals