Members-Only
Recent Talks & Demos are for members only
You must be an AI Tinkerers active member to view these talks and demos.
DoneThat: Preventing Prompt Injection
This talk covers an MVP solution for preventing prompt injection in an AI work tracker, focusing on LLM-based flagging, implementation choices, and a live demo.
DoneThat is an AI work tracker - it takes screenshots and analyzes those to automate time tracking. It has several opportunities for prompt injection:
- A chat-based agent
- “Memories” that refine processing
- The screenshots themselves
I will walk through the basics for preventing abuse (rate limiting, authentication, etc) but then focus on preventing prompt injection. I am working on this right now so can’t say what the exact solution will be but most likely using LLMs to identify and flag potential abuse. I will walk through the options considered, why I selected the one I chose, and the final implementation on live data. Can also do a live demo.