liveBook · Manning

Data exfiltration

As explained in AI Governance

Data exfiltration is the unauthorized or inappropriate transfer of data from an internal system to an external destination.

Data exfiltration is the unauthorized or inappropriate transfer of internal data to an external destination. In agentic systems, outbound tool calls to APIs or other services can create this risk; runtime tool discovery can also expose agents to unvetted tools. section 5.6.1

Reduce exposure by enforcing purpose limits at the system boundary:

  • Use purpose-locks to constrain the agent’s design and tool permissions. section 5.6.1
  • Route every outbound tool call through a purpose-aware egress gate. OPA can decide whether a call is permitted, but a separate enforcement point must intercept or block it. section 5.6.1

How AI Governance covers it

How deep it goes
Traces tool-call and runtime-discovery risks, then outlines control architecture without detailing policy rules.
Examples
Uses a scheduling agent that infers diabetes from appointment history and adds a sensitive profile tag to illustrate purpose creep.
What you can do after
Apply purpose-locks and a purpose-aware egress gate, using OPA for decisions and a separate enforcement point to intercept tool calls.
How it connects
Connects data exfiltration to purpose limitation, data minimization, tool permissions, dynamic tool discovery, and subprocessors.

Also explained in

Compare how 3 books explain data exfiltration →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of AI Governance. The section links point to the parts of the book it draws on.