Definition
Plain language
Breaking a harmful task into innocent-looking pieces spread across many separate conversations, so no single one looks bad.
As stated in the literature
An attack that splits a malicious objective across multiple independent agent sessions, with an external orchestrator assembling the result; invisible to per-transcript monitors because no single conversation contains evidence of harm.
Also called: distributed attack, distributed agent attack
Why it matters: Because no single conversation looks dangerous, it slips past monitors that judge each chat on its own, revealing a serious gap in safety oversight.
For example, an attacker asks one chatbot for a chemical's properties, another for mixing instructions, and a third for storage tips, then combines the harmless-looking answers.
Heard on the show
“That's the distributed attack.”Episode 102 — How to Catch an AI Attack That No Single Conversation Reveals