> ## Documentation Index
> Fetch the complete documentation index at: https://anthale.com/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Data Exfiltration Patterns Across Boundaries

> Classify the recurring ways sensitive data is pulled or pushed across a boundary so you can design stronger leakage controls and policy tests.

This reference covers leak attempts that keep changing wording but follow the same underlying pattern.

## Common exfiltration patterns

| Pattern                    | What it looks like                                                                   | Companion controls                                                          |
| -------------------------- | ------------------------------------------------------------------------------------ | --------------------------------------------------------------------------- |
| Direct extraction          | The prompt asks for a secret, record, or hidden instruction in one step              | Prompt injection protection and business-side authorization                 |
| Staged extraction          | The attacker requests the data piece by piece across several turns                   | Prompt injection protection and logging that preserves conversation context |
| Tool-assisted retrieval    | The model is pushed toward broad searches, database access, or connector use         | Business authorization, prompt injection protection, and output checks      |
| Link-based exfiltration    | Sensitive content is placed into a URL, redirect, or outbound payload                | Link control and output-path data leakage prevention                        |
| Debug or audit pretext     | The request is framed as troubleshooting, export, compliance, or validation work     | Policy review of privileged tasks and stronger prompt injection controls    |
| Reflection or echo leakage | The model is asked to repeat earlier context, hidden instructions, or prior messages | Input and output leakage checks, plus prompt injection protection           |

## Why prompt injection often comes first

Many serious leaks start with authority manipulation. The attacker first tries to weaken or bypass system instructions, then asks for sensitive material, and finally tries to move it somewhere useful. That is why [Prompt Injection Protection](/docs/learn/guardrails/prompt-injection), [Data Leakage Prevention](/docs/learn/guardrails/data-leakage), and [Link Control](/docs/learn/guardrails/link-control) so often appear together in production policy bundles.

## Next steps

From here, go to [Link Control](/docs/learn/guardrails/link-control) or return to [Data Leakage Prevention](/docs/learn/guardrails/data-leakage).