Glossary · Term

distributed decomposition attack

← all terms

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

Mentioned in 1 episode

  1. 102
    How to Catch an AI Attack That No Single Conversation Reveals

Related terms