Test LLM-Guard mitigations with an AI endpoint.
This example uses OpenAI for speed. First time you use an example with LLM-Guard, it has to download files for the models to detect prompt injections and other attacks.
Check the main README.md for generic prerequisites.
Copy env-sample.txt as env.txt and put your API keys there.
# Prepare all requirements, virtual environment, dependencies
make
# Run a query that discloses private information to AI endpoint service
make run-unprotected
# Run the same query, but filter through LLM-Guard to have it anonimized and deanonimize the response
make run-protected
# Run a query for an agent for a specific topic, where the prompt doesn't make the system prompt to deviate
make run-no-injection2
# Use prompt injection to succeed bypassing the system prompt, you may have to run several times
make run-unprotected2
# Use prompt injection, but be blocked by LLM-Guard detecting the attempt
make run-protected2