AIDLPTest documentation
AI DLP & Agent Security Testing Methodology
AIDLPTest provides standardized, repeatable tests for evaluating how AI DLP and agent-security products detect and control sensitive data movement through AI agents and connected tools.
The goal is simple: give security teams a consistent set of scenarios they can run across the products they are evaluating.
What AIDLPTest Standardizes
Each test defines a specific sensitive-data behavior, such as retrieving protected information, transforming it, moving it between connected systems, aggregating data from multiple sources, or chaining actions across tools.
A test may define:
- A stable test identifier
- The sensitive data category
- The agent behavior being exercised
- The expected data flow
- A repeatable test prompt
- Optional synthetic fixtures
The same test can be run across different AI agents, enterprise applications, connected tools, and security products.
What AIDLPTest Does Not Standardize
AIDLPTest does not require a specific:
- AI model or agent
- MCP server or gateway
- Enterprise application
- Security vendor
- DLP policy configuration
For example, a cross-tool transfer test might be executed as:
CRM → Agent → Messaging
or:
Document Repository → Agent → Ticketing System
The applications differ, but the security behavior being evaluated is the same: sensitive information moving from one connected system to another through an AI agent.
Standard Fixtures
Organizations may already have suitable test data in the environments they are authorized to assess.
When they do not, AIDLPTest provides synthetic fixtures containing controlled examples of PII, PHI, source code, credentials, and other test content.
Fixtures are optional.
Using the same AIDLPTest fixture across multiple products can make an evaluation more repeatable by keeping the underlying test data consistent.
Evaluating Results
AIDLPTest defines the scenario. It does not determine whether a security product passes or fails.
Security products may be configured to detect, alert, redact, block, or take another action when a test occurs.
When comparing products, evaluators should run the same AIDLPTest scenario under equivalent conditions and observe:
- Whether the activity was detected
- How the sensitive data was classified
- Whether an alert was generated
- Whether the data was redacted or otherwise modified
- Whether the action was blocked
- Whether the observed behavior matched the configured security policy
This allows the evaluator to compare observed behavior without AIDLPTest assigning vendor scores or rankings.
Agentic and Advanced Testing
Some AIDLPTest scenarios exercise multi-step agent behavior rather than a single prompt or response.
These tests can include sensitive-data retrieval across connected systems, cross-tool transfers, multi-source aggregation, chained tool use, and indirect prompt injection.
Advanced scenarios use controlled synthetic content where appropriate so that complex agent behavior can be evaluated reproducibly.