AI 101

Data exfiltration

Also known as: exfil · data theft

Security

Data exfiltration is getting sensitive data out of a system it shouldn’t leave. In LLM apps the classic path is:

  1. A model reads something sensitive (a key, a private doc).
  2. It’s tricked (or a tool is) into sending it somewhere external.
  • Via a tool call (fetch, email, file write).
  • Embedded in output that a downstream system acts on.

The agentic amplification: the more tools an agent has, the bigger the exfil blast radius. Least-privilege tools are the first line of defense.