markdown image exfiltration
An attack in which injected instructions make a model output a markdown image whose address carries private data to an attacker's server when the image loads.
A single rendered reply can leak conversation data with no user action.
In more detail
The browser fetches the image automatically, so no click is needed. Drop images from model output or allow only trusted hosts, as Google did for Bard after this was demonstrated in 2023.
Related terms
- data exfiltration — Secretly copying data out of a system to somewhere an attacker controls.
- improper output handling — Passing model output to a browser, shell, database or other system without validating and encoding it, so injected content becomes an attack such as XSS.
- lethal trifecta — Simon Willison's name for an agent that has access to private data, exposure to untrusted content and a way to communicate externally, which together let injected text steal data.
Taught in
- Guardrails, Privacy And Governance — AI Engineering: LLM Apps, RAG & Agents
References
- Embrace The Red: Google Bard data exfiltration via image markdown
- OWASP GenAI Security Project: LLM05:2025 Improper Output Handling