Insecure Agentic API Examples

This repository contains small, intentionally insecure examples that demonstrate common failure modes in agentic or tool-using LLM APIs. Each script focuses on a specific vulnerability pattern to make it easier to study, reproduce, and harden against in production systems.

Contents

• parameter_tampering.py: A full chat-completions wrapper that forwards arbitrary request parameters to the OpenAI client.
• raw_sti.py: A raw tool-call injection example that executes SQL using untrusted tool-call arguments.
• role_injection.py: A multi-agent runner that exposes code generation and execution flow to prompt/role injection.
• unbounded_consumption.py: A simplified chat API with tool execution that can trigger unbounded tool usage or resource consumption.

Requirements

• Python 3.10+
• An OpenAI API key (for scripts that call the OpenAI client)
• Optional: PyTorch and Transformers (for raw_sti.py)

Suggested dependencies (adapt as needed for your environment):

pip install fastapi uvicorn openai httpx pydantic
pip install torch transformers  # for raw_sti.py

Set your API key before running the FastAPI examples:

export OPENAI_API_KEY="your-key"

Running the FastAPI Examples

Each FastAPI script defines its own app. Use uvicorn to run them locally:

uvicorn parameter_tampering:app --reload
uvicorn role_injection:app --reload
uvicorn unbounded_consumption:app --reload

Then send requests to the exposed endpoints (e.g., /chat or /ask) with your preferred HTTP client.

Blogs

For more background on the vulnerability patterns showcased here, check out the following write-ups.

Learn more at:

• Special Token Injection (STI) Attack Guide