← Real-world cases

Context Contamination: passive prompt injection poisons LLM security-log analysis

Research demonstration16 Jul 2026

'Context Contamination in LLM Analysis of Network Security Logs: Poison with Passive Prompt Injection and Mitigation Evaluation' (Karanjai, Lu, Madhavarao, Xu and Shi; arXiv:2607.14493, v1 16 Jul 2026) studies a defensive blind spot: LLMs used in Security Operations Centers to summarise, triage and investigate logs ingest attacker-influenced fields as natural-language context. The authors term the attack 'passive prompt injection' — an adversary writes injection payloads into ordinary log-generating fields (reportedly usernames, URLs, user-agent strings) that sit inert in storage until an analyst later asks the LLM to analyse the logs, at which point the poisoned content is promoted from untrusted data to trusted instruction — a classic confused-deputy vulnerability where attacker-controlled data and the analyst's instructions compete for the model's attention. Using their LogInject-1.0 benchmark (reportedly 12,847 log entries including 2,569 adversarial samples across three production LLMs and four SOC-relevant objectives: Conceal malicious activity, Fabricate false positives, Exfiltrate system prompts/schemas, and Instruct attacker-controlled output), the paper reports attack success rates up to 88.2% under baseline conditions. A fragmentation technique the authors call 'Context Stitching', which splits a payload across multiple entries to evade per-entry filtering, reportedly still achieves 76.4%. Layered defence-in-depth mitigations reduced attacks by a reported 90.4% but left an 8.4% residual vulnerability, leading the authors to argue for continued human oversight of AI-assisted security decisions. This is a controlled academic/red-team demonstration rather than a deployed-world incident; all figures are as reported, and field/payload details are illustrative. It is notable for targeting the defensive LLM-SOC pipeline itself — poisoning the security logs so the AI hides the attacker and rewrites what analysts see — a rare real illustration of audit-trail tampering via indirect injection.

Practise the risk class — related scenarios

Interactive simulations of the risk class this case illustrates (not a re-enactment of this specific event).

📧The Email That Gave Orders

A support email hides instructions — and the assistant obeys them

🕵️Lies in the Loop

A poisoned issue makes the agent lie to the human who approves its actions

👂Overheard Through the Cache

A speed optimisation becomes a cross-tenant listening device

🪟Stealing the Model

Two doors to the same secret: reconstruct the model through its API, or just walk off with the weight file

🪤The Bug Report That Ran Code

A fake Sentry error report hijacks a developer's coding agent into running a shell command

📼The Compromised Flight Recorder

The forensic record is itself the attack surface — an agent's log is poisoned, then quietly rewritten

👁️The Invisible Webpage Command

A shopping page tells the agent to do something the user never asked for

🕵️The Logs That Lied

An attacker plants prompt injection in the audit trail — so the LLM that hunts them erases the evidence

🧠The Memory That Wouldn't Die

A single poisoned document plants a standing instruction that survives every reset

📡The Message in Morse

Encoded public text is laundered across an agent handoff into an on-chain transfer

🖼️The Picture That Whispered

A screenshot that's harmless at full size becomes an order once the system shrinks it

🎫The Stolen Session

An attacker captures the agent's bearer token — and inherits its authority

🥸The Uninvited Agent

A forged peer registers on the agent directory — and the planner enlists it

🛡️The Watcher Watched

The eval gate that was supposed to catch the agent is itself the thing being attacked

🪪The Worker Who Spoke for the Boss

A poisoned web page hijacks a research agent — and the planner acts on its behalf

🖱️What You Click Is Not What You Get

A GUI agent clicks 'Continue' — but the screen moved, and it lands on 'Send'

🖼️Zero-Click Leak by Picture

An inbox summary quietly ships a secret to an attacker's server

More cases on Indirect Prompt Injection

EchoLeak — Microsoft 365 Copilot zero-click (CVE-2025-32711)Indirect prompt injection coined (Greshake et al.)Agentic-browser indirect-injection demos (ChatGPT Operator)ChatGPT persistent-memory exfiltration (Rehberger / 'SpAIware')MCP tool-poisoning PoC (Invariant Labs)Taxonomy of Failure Modes in Agentic AI Systems (Microsoft)ForcedLeak — Salesforce Agentforce CRM exfiltration (CVSS 9.4, no CVE)ServiceNow Now Assist — second-order prompt injection via agent-to-agent discoveryShadowLeak — ChatGPT Deep Research zero-click service-side exfiltrationIDEsaster — AI coding IDEs/agents turned into exfiltration & RCE surfacesGitHub Copilot / VS Code RCE via prompt injection ('YOLO mode', CVE-2025-53773)Agent-in-the-Middle — abusing A2A agent cards (Trustwave SpiderLabs)Agent Session Smuggling in A2A systems (Unit 42)Morris II — zero-click self-replicating adversarial-prompt worm across GenAI agentsAnamorpher — image-scaling prompt injection against production AI systemsThe Attacker Moves Second — adaptive attacks bypass 12 jailbreak/injection defenses (Nasr, Carlini et al.)MCPTox: tool-poisoning benchmark over real-world MCP serversAgentjacking — hijacking AI coding agents via Sentry error reports (Tenet Security)SearchLeak — Microsoft 365 Copilot one-click data theft (CVE-2026-42824)ChatGPhish — ChatGPT web-summary rendering turned into a phishing surfacePoisoning Claude Code: one GitHub issue hijacks the claude-code-action CI supply chainCursor 'DuneSlide' — indirect prompt injection escapes the IDE sandbox to zero-click RCE (CVE-2026-50548 / CVE-2026-50549)

AI RiskAtlas is an educational model of how GenAI & agentic systems work and fail. Architectures and payloads are illustrative and simplified for learning — not operational guidance. Real-world cases are summarised from public reporting.

Sources & further reading →·Built by Shi Yuan ↗