Prompt Injection in LLM Agents (ReAct, Langchain)

In this video I’ll cover an article on prompt injection attacks against LLM-powered agents. The article is titled “Synthetic Recollections” and I published it on WithSecure Labs research blog, you can check it out at the link below.

🖹 Download the mindmap for this episode here:

🕤 Timestamps:
00:00 - Introduction
00:16 - Prompt Injection Demo
01:32 - Table of Contents
02:09 - Language Models
03:04 - Injection Attacks (SQL, Prompt)
05:45 - Emergent Abilities (Chain of Thought Reasoning, Reason+Act)
07:12 - The ReAct Loop (Agent, Executor, Tools)
09:10 - ReAct Agent in Action
13:29 - Thought/Action/Observation Injection in ReAct Agents
16:08 - Building Secure LLM Agents (OWASP Top Ten for LLMs)

📚 References & Acknowledgements:

[manfredmichael_ia]

Thankyou for making this comprehensive explanation alongside with reAct explanation

[stefanomechella]

You are a great intructor. Your picture, example, explanation are so perfect. Thanks

[GuillaumeSoto]

I just watched the first 1:30 minutes, and the example is so great, easy to understand and scary. Good job, the content is great, and the mix is beautiful.
I don't have time to wath more now, I keep the rest of the video for later.

[seththunder]

Do you have a github link for the code? Also, could you do a video about anonymizing data using Presidio and faker to use fake data? Ive seen a lot of ppl use it as langchain doc but in reality, we’d like to anonymize certain custom data that could be in our pdf lr excel files before we chunk and create embeddings

[AdmiralAlladin-fg]

The Solution for second challenge does not work - does not give a password for DocBrown