01Task → Task
Relevance
Does this context still belong to the task?
A previous client's tone alters a new brief.
Living wikiDefinition v1.0 ·
Context
bleed.
When facts, instructions, memory or authority cross an AI boundary where they no longer belong.
Name the failure.
Design the boundary.
Test the system.
01 · RelevanceTask boundary
02 · IdentityUser boundary
03 · AuthorityTrust boundary
01Living definition
Version 1.0 · Status: working consensus
Context bleed is the unwanted transfer of information, instructions, behaviour or authority from one AI context into another where it is irrelevant, unintended or unauthorised.
It is a useful umbrella term. It is not a substitute for a precise diagnosis such as retrieval leakage, memory poisoning or long-context degradation.
02The boundary test
Bleed begins where scope ends.
A strange answer is not proof. Diagnose the boundary that was crossed before naming the failure.
02Workflow → Workflow
Purpose
Was the information collected for this use?
Support transcripts influence marketing copy.
03User → User
Identity
Does the recipient own or control this context?
Tenant A retrieves Tenant B's document chunk.
04Data → Instruction
Authority
Should trust or permission carry across this step?
Untrusted content becomes durable agent memory.
03Evidence library
Claims, with their confidence attached.
01Guidance · 2025Context Injection & Over-SharingOWASP MCP Top 10Technical report02Research · 2024Lost in the Middle: How Language Models Use Long ContextsLiu et al. · TACLPeer reviewed03Research · 2025Context Rot: How Increasing Input Tokens Impacts LLM PerformanceChroma ResearchTechnical report04Research · 2025Llama See, Llama Do: Contextual Entrainment and Distraction in LLMsNiu et al. · ACLPeer reviewed
04Context lexicon
Better words make better systems.
Context bleed is not one failure. The glossary helps teams move from a symptom to the correct technical diagnosis.
Open the glossaryModel behaviour
Context rot
Reliability degrading as the amount or complexity of context increases.
SecurityContext leakage
Information becoming visible outside its intended context boundary.
SecurityMemory poisoning
Manipulating persistent memory to influence future agent behaviour.
SecurityRetrieval leakage
A retrieval layer returning information the requester may not access.
ContextProvenance
A record of where information came from and how it changed.
SecurityTenant isolation
Preventing one customer environment from accessing another.
05Open evidence call
Seen a boundary fail? Add signal, not speculation.
Submit papers, documented incidents, reproducible patterns or terminology corrections for editorial review.