Guidance · 2025Technical report
Context Injection & Over-Sharing
OWASP MCP Top 10
Security guidance for contexts that are shared, retained or insufficiently scoped across users, agents and workflows.
MCPcross-userisolation
Open evidence noteEvidence library
Evidence before narrative.
Academic research, security guidance, technical reports and observed patterns — labelled so a preprint never masquerades as a proven incident.
- Records
- 08
- Last review
- 20.07.26
Showing 8 of 8 records
Research · 2024Peer reviewed
Lost in the Middle: How Language Models Use Long Contexts
Liu et al. · TACL
Shows that language models often use relevant information less reliably when it is positioned in the middle of long inputs.
long contextpositional biasevaluation
Open evidence noteResearch · 2025Technical report
Context Rot: How Increasing Input Tokens Impacts LLM Performance
Chroma Research
Evaluates 18 models and documents non-uniform performance degradation as input length increases.
context rotbenchmarkslong context
Open evidence noteResearch · 2025Peer reviewed
Llama See, Llama Do: Contextual Entrainment and Distraction in LLMs
Niu et al. · ACL
Examines a mechanism through which irrelevant tokens already present in context pull later model generations towards them.
distractionmechanisticirrelevant context
Open evidence noteResearch · 2026Preprint
Instruction Bleed: Cross-Module Interference in Prompt-Composed Agentic Systems
Lin & Liu · arXiv
Proposes compositional behavioural leakage as a measurable form of interference between prompt modules sharing a context window.
agentsinstructionsinterference
Open evidence noteResearch · 2026Preprint
From Untrusted Input to Trusted Memory
Dash et al. · arXiv
Studies how a single adversarial memory write can retain influence across later agent interactions.
memorypoisoningagents
Open evidence noteCase pattern · 2026Technical report
Multi-tenant retrieval leakage
Recurring architecture pattern
A shared vector index returns chunks belonging to another tenant because access policy is absent or applied after retrieval.
RAGtenant isolationaccess control
Open evidence noteField report · 2026Community evidence
Cross-project context bleed and directory drift
Google AI Developers Forum
A community bug report describes an IDE agent reading or acting within the wrong project directory.
coding agentsprojectscommunity report
Open evidence note