AI Data Exfiltration

What AI Data Exfiltration is

Exfiltration can occur through crafted prompts, indirect injection that instructs an agent to send data outward, over-broad retrieval permissions, or rendering that triggers an outbound request.

How it works

Defences include per-user permission filtering at retrieval time, blocking outbound network capability by default, sanitising rendered content, and logging every tool call for review.

Why it matters

Retrieval systems frequently inherit an index that ignores the user's own access rights, which turns a helpful assistant into a data leak.

Common uses

• →Security review of RAG deployments
• →Agent network policy design