Data Exfiltration
Data exfiltration is the unauthorized transfer of data from an organization's systems to an external location. In AI agent contexts, exfiltration risk arises when an agent with access to sensitive data also has the ability to communicate externally, creating a channel through which data can be extracted.
Why It Matters
AI agents create novel exfiltration paths. An agent with read access to a customer database and write access to email can copy customer records into an outbound message. An agent with access to internal documents and a web browsing tool can send document contents to an external endpoint.
Compositional risk is the exfiltration enabler: no single permission creates the risk, but the combination of internal read access plus external write access creates the channel. Per-action policies that restrict external tool calls when the session already holds sensitive data access are the primary mitigation.
Credential mediation provides a secondary defense: if the agent's context is compromised through prompt injection, the attacker cannot extract credentials because the agent never held them.