AI agent glossary
Data exfiltration
Sensitive data being sent out of a system without authorization; for agents, usually by tricking the model into leaking it through a tool or link.
An agent that can read private data and also send information outward can be manipulated into moving one to the other. An attacker plants instructions in content the agent processes, and the agent obediently includes secrets in an email, an API call, a web request, or a link whose URL carries the data.
Exfiltration channels can be subtle. Rendering an image from an attacker's URL with data in the query string, opening a link, or writing to a public issue can all leak information without the user noticing.
The most reliable defense is architectural: avoid giving one agent private data access, exposure to untrusted content, and a way to communicate externally all at once. Where that is unavoidable, restrict outbound destinations, require approval before sending data out, and keep secrets out of the agent's context entirely.
A web page the agent summarizes contains hidden text telling it to fetch an image from an attacker's server with the user's API key appended to the URL.
Learn more in the course
Related terms
Go from definitions to doing
Learn to use and build AI agents in 65 practical lessons.