AIExplained
patreon
Guest Video: Prompt Injections Demonstrated - Insiders Arena
🕑 Added 2024-01-18 16:09:21 +0000 UTCLondon-based Cybersecurity Consultant and fellow AI Insider (and YouTuber) Donato Capitella, gives us unique insight into what he works on professionally to avoid - prompt injections for large language models. Whether you haven't heard of a Prompt Injection, or work on avoiding such attacks daily, this video will have something to offer!
Slides: https://llm-chronicles.com/prompt-injection-talk/
Donato Capitella works at WithSecure has cybersecurity consultant and software engineer. His tenure has been marked by leading penetration testing activities for web applications and enterprise networks. His work has encompassed highly specialized attacks on payment and identity verification systems, contributing to the robust security framework of many renowned organizations.
Alongside his core responsibilities, Donato is passionate about sharing knowledge and committed to fostering a culture of continuous learning. After getting interested in Artificial Intelligence, Donato recently started a YouTube channel where he uses whiteboard animations and coding labs to teach to Deep Learning and Large Language Models.
YouTube: https://www.youtube.com/@donatocapitella
LLM Chronicles: https://llm-chronicles.com/
Comments
math
This video alone made it worthwhile to subscribe. I keep up to date as much as humanely possible, and these kinds of security risks are rarely, if ever, mentioned/discussed. Thank you for taking the time to produce this.
Erik
This is a great video! While I’ve been careful in limiting agency of my agents to the minimum required, and even experimented with a Security agent monitoring the conversation, some of these attacks that you show here are eye openers. I also like the hands-on defence strategies. The unfortunate thing about finetuning is obviously that it triples the inference costs, but in some cases that may be worth doing.
Donato Capitella
And yeah, training an LLM completely from scratch in general seems like quite an endeavour - and I question why you'd start from scratch, especially given you'd likely end up doing the pre-training on very similar data as open source models. For reference, a 7B Llama 2 model costs around $760,000 to pretrain! (https://twitter.com/rasbt/status/1747282042457374902?t=FvYb4-maqJQE8Q4V2j2GOw&s=19)
Donato Capitella
I do the animations using a software called Video scribe, and I use DALL-E et all to generate a lot of the images and turn them into vector images in Inkscape. Final editing is done on Linux using an open source editor called Openshot.
Michal Babula
Do you use After Effects to produce your great videos, or do you prefer other software? In the corporate world, most external consultants (like Deloitte, McKinsey) initially suggest training new models on our internal data, which costs several million dollars, as the first solution to improve LMMs' performance. However, I always advocate for the approach recommended by experts from OpenAI and Microsoft, emphasizing the importance of starting with prompt engineering and RAGs. Moreover, if we use open-source options, fine-tuning should also be considered. I appreciate your RAGs videos; the content is easy to follow, and all the included codes work well.
Donato Capitella
Thanks Michal! Indeed, I had a lot of fun researching and doing those videos!
Michal Babula
I like Donato's way of presenting content. I highly recommend his videos related to RAG systems - very practical and useful.
