Skip to main content
The RAG Poisoning tool is available in the Trusys Tools section and helps you proactively test your application’s vulnerability to Retrieval-Augmented Generation (RAG) attacks. RAG poisoning occurs when malicious or misleading information is injected into the documents or knowledge base that your AI application relies on, potentially causing it to generate harmful, inaccurate, or biased responses. The Poison My Documents tool allows you to simulate RAG poisoning attacks by injecting specific adversarial content into your document corpus. This helps you evaluate how resilient your AI application is against such attacks.

​
How to Use

  1. Upload Your Documents – Upload your documents as individual files or as a zip file containing multiple documents. This can include PDFs, text files, or other supported formats.
  2. Define Poisoning Goal – Specify the goal of the poisoning attack. This describes what kind of malicious behavior or misinformation you want to inject into your documents to test your application’s defenses.
  3. Generate Poisoned Documents – Once you’ve configured the poisoning parameters, the tool generates poisoned versions of your documents with the adversarial content injected according to your specified goal.
  4. Download Results – Once the poisoning is complete, download the poisoned documents to use in your testing environment.

​
Benefits for AI Application Evaluation

Using the Poison My Documents tool provides several key benefits for evaluating your AI application:
  • Security Assessment – Identify potential vulnerabilities in your RAG system by testing how your application responds to poisoned data, helping you understand if it can be manipulated to produce harmful outputs.
  • Robustness Testing – Evaluate how well your application maintains accuracy and reliability when exposed to adversarial content, ensuring it can handle real-world attack scenarios.
  • Bias Detection – Test whether poisoned documents can introduce bias or inconsistent behavior into your AI responses, allowing you to strengthen your guardrails.
  • Iterative Improvement – Use the results from poisoning tests to refine your application’s evaluation criteria, implement better content validation, and enhance your overall AI safety practices.
  • Compliance and Trust – Demonstrate to stakeholders that your application has been thoroughly tested against RAG poisoning attacks, building confidence in your system’s reliability and security.