Skip to content
CBContext/Bleed Wiki
← Evidence library
Preprint

Research · 2026

Instruction Bleed: Cross-Module Interference in Prompt-Composed Agentic Systems

Lin & Liu · arXiv

What it says

Proposes compositional behavioural leakage as a measurable form of interference between prompt modules sharing a context window.

Why it matters here

An early attempt to formalise ‘bleed’ language for prompt-composed agent systems; findings remain provisional.

Editorial caution

This record has not completed peer review. Treat terminology, measurements and generalisations as provisional.

agentsinstructionsinterference
Open original source