AI Canary Tokens

What Are AI Canary Tokens?

AI Canary Tokens are traceable fake data that you plant in your documents, code, databases, or training data. If an AI model ever outputs these tokens, you have proof that your data was accessed without authorization.

Think of them as "digital bait." Just like a canary in a coal mine alerts you to danger, an AI Canary Token alerts you when your sensitive data has leaked into an AI system.

Unlike traditional DLP that only catches data at the point of entry, AI Canary Tokens detect leakage at the point of output - when an AI model regurgitates your protected information.

How AI Canary Tokens Work

1 Generate Unique Canaries

Create fake but realistic-looking data: AWS credentials, API keys, SSNs, internal project details, or custom markers. Each canary is unique to your organization.

2 Plant in Your Data

Embed canaries in strategic locations: internal documentation, code repos, configuration files, training datasets, or any data that might be scraped for AI training.

3 Monitor AI Outputs

Use our SDK or API to scan AI-generated content. When a canary appears in an AI's output, you get instant notification with full provenance.

4 Prove Leakage

Because each canary is unique and traceable, you can prove exactly which document was accessed, when it was leaked, and through which AI application.

Types of AI Canary Tokens

Signal Canary supports multiple classes of canaries, each with different detection characteristics:

Class	Description	Best For
Exact	Cryptographic markers like `CANARY-A1B2C3D4...`	High-confidence detection, code comments, config files
Structured	Realistic fake credentials: AWS keys, API tokens, SSNs, credit cards	Training data, internal docs, detecting credential leakage
Semantic	Unique facts like "Project ALPHA budget is $450,000"	Detecting paraphrased content, RAG systems, summaries
Active	Fake API endpoints and webhook URLs that report when accessed	Detecting AI agents, automated systems, code execution

Semantic canaries are especially powerful because they survive paraphrasing. Even if an AI rephrases "Project ALPHA budget is $450K" as "the ALPHA project has approximately half a million in funding," we can still detect it.

Why AI Canary Tokens Matter

Traditional DLP stops data at entry. AI Canary Tokens prove leakage at output. This is critical because:

Training data is often scraped - You may not know when your data entered an AI model
RAG systems pull from many sources - Your documents may be indexed without your knowledge
Paraphrasing evades traditional detection - AI models reword content, but semantic canaries persist
You need proof for legal action - Canary provenance provides evidence of unauthorized access

Without AI Canary Tokens, you have no way to prove that an AI model accessed your proprietary data. With them, you have a traceable audit trail.

Getting Started

Ready to protect your data? Here's how to begin:

Go to AI DLP > Canary Tokens in your dashboard
Create your first token - choose a type (API key, SSN, etc.)
Copy the canary value and plant it in your documents
Set up monitoring via our SDK or API

For automated detection across your AI applications, see our RAG Testing Guide.