App Monitoring
Exfiltration Link
Also: markdown image exfiltration
A URL in model output that carries stolen data in its address, sent to the attacker when a client loads or the user clicks it.
In more detail
An injected prompt can make a model write an image or link such as . When the chat client renders the image it fetches the URL, delivering the data without any click. Defenses include blocking unknown domains in rendered output and stripping query strings from model-written links.