Prompt Injection: When Hackers Befriend Your AI - Vetle Hjelle - NDC Security 2024

This talk was recorded at NDC Security in Oslo, Norway. #ndcsecurity #ndcconferences #security #developer #softwaredeveloper
Attend the next NDC conference near you:
ndcconferences...
ndc-security.com/
Subscribe to our KZbin channel and learn every day:
/‪@NDC‬
Follow our Social Media!
/ ndcconferences
/ ndc_conferences
/ ndc_conferences
This is a technical presentation where we'll look at attacks on implementations of Large Language Models (LLMs) used for chatbots, sentiment analysis, and similar applications. Serious prompt injection vulnerabilities can be used by adversaries to completely weaponize your AI against your users.
We will look at how so-called "prompt injection" attacks occur, why they work, different variations like direct and indirect injections, and then see if we can find good solutions on how to mitigate those risks. We'll also learn how LLMs are "jailbroken" to ignore their alignment and produce dangerous content.
LLMs are not brand new, but we know that their use will increase drastically in the next few years, and therefore it is important to take security seriously by considering the risks involved before using AI for sensitive operations.

[@ManuelBasiri]

So basically, unless your data is already open and public, don't give it to LLMs otherwise, it will become open and public.

[@monad_tcp]

46:34 doesn't fully work, well it worked for OpenAI itself, but they have much more resources than everyone else

[@Roibarkan]

17:23 An explainer about word embeddings: kzbin.info/www/bejne/nYLHlaeKmdJ6lZo

[@goldnutter412]

Yeah great video en.wikipedia.org/wiki/Stochastic_parrot though 🤣 whoever came up with this is a legend🥰

[@goldnutter412]

16:16 kzbin.info/www/bejne/oZ2qiKmIqLGEgbc 🤣