← Back to concept explainers

Concept Fable

Data exfiltration

Conclusion first:Data exfiltration is unauthorized movement of data outside a controlled boundary. In Agent systems, it can happen through tool calls, prompts, logs, browser actions, API requests, file uploads, summaries, screenshots, vector retrieval, or malicious instructions. Controls include egress filtering, DLP, scoped context, redaction, approvals, audit logs, and tenant-aware policies.

Data Exfiltration · Security

Read the story first. There is no rush to give a definition here; let the question grow on its own first.

Fable

The thin threads flowing out from the accounting house

The lord's manor had a secret vault holding tax ledgers, contracts and defense maps. The guards feared thieves picking the locks most of all, so they added three iron bolts to the door, and everyone entering or leaving had to be logged.

For a long time, the secret vault was safe. Until one day, a merchant from outside the city accurately stated the stock levels of several granaries. The locks were not broken, the iron bolts showed no pry marks, and the register showed no stranger had entered.

The guards thought someone had slipped in at night, so they added more patrols. But the information still leaked out slowly: first the trade names in the tax ledgers, then the discounts on the contracts. What leaked each time was never the whole book, just a few seemingly harmless sentences.

Later, a copyist was found. He had not carried the ledger out; he had merely slipped library numbers into replies to foreign merchants. Sometimes he used "examples" to reveal real prices, and sometimes he pieced several fragments of information into a code.

The first fix by the prefecture was to forbid scribes from taking paper out of the building. But he could still memorize the contents or hide them in ordinary letters. Everyone then realized that secrets leaving the vault do not necessarily look like stealing a box; they can also seep away like water through a fine crack.

The new steward moved the line of defense to the exit. He decreed that any outgoing letter must be checked for details from the secret vault; scribes could see only the small portion needed to complete the task; and unusual inquiries must be paused and reviewed by a designated person.

The new steward also re-examined the scribe’s duties. In the past, he could see the entire tax ledger just to reply to a very small letter; now each task only receives the corresponding pages, and temporarily requesting more materials requires justification. How much one can access is tied to what must be completed. The steward also set up bait questions. If a foreign merchant asked several seemingly unrelated small questions in succession that together could piece together grain warehouse reserves, the letter would be intercepted. In the past everyone only guarded against the whole ledger being stolen; now they began guarding against fragments being taken away bit by bit. The office also began recording every intercepted letter. Which were false interceptions, which were real leaks, which exposed overly broad permissions—all had to feed back into changing the rules. The steward said leaks are often not a loud bang, but a series of seemingly normal small exits.

A few months later, there were far fewer "casual slips" in the prefecture. The guards still watched the gates, but they no longer focused only on the iron bolts. The real danger was that people or things allowed access to secrets could package secrets as normal communication and carry them out.

Reveal

This story is about: data leakage

Data exfiltration is unauthorized movement of data outside a controlled boundary. In Agent systems, it can happen through tool calls, prompts, logs, browser actions, API requests, file uploads, summaries, screenshots, vector retrieval, or malicious instructions. Controls include egress filtering, DLP, scoped context, redaction, approvals, audit logs, and tenant-aware policies.

What matters is that it is not just a term, but a structural problem that repeatedly appears in real AI / Agent systems. Understanding it lets you judge when to add a model, when to change a process, and when to add governance.

Metaphor mapping

  • Treasury iron gate: traditional perimeter security
  • Many thin lines: small-scale exfiltration through legitimate tools and outputs
  • External translation/analysis tools: third-party APIs or cloud services
  • Governance by egress: egress control and DLP
  • Sensitive field tagging: de-identification and context minimization
  • Final insight: Agent data security must manage where data goes, not just who can log in

Andy Ren Xu's judgment

The practical value of this concept is that it helps you break "AI features that look smart" into deliverable, verifiable, and governable units of work.

data leakageprompt injectionegress controlDLPaudit logs

Related distinctions

This concept is easily confused with adjacent concepts; it is recommended to continue to the comparison page.

京ICP备2026014272号