[Webinar] Building LLM applications in a secure way (WithSecure™)

This is the recording of a webinar we did at WithSecure about the risks of creating LLM applications that act as autonomous agents, and what can be done to mitigate these risks.
Info and security canvas download: www.withsecure....
Timestamps:
00:10 - Where did LLMs come from?
05:36 - Building LLM applications
06:52 - LLM agents
08:10 - Misconceptions about AI safety
09:58 - Risks of LLM use-cases
11:30 - Prompt injection (demo)
18:46 - LLM agents
24:00 - Prompt Injection Demo in Browser Agent (Taxi AI)
30:06 - Root cause of LLM alignement issues
34:45 - Comparison with traditional injection attacks
37:14 - Controls and defences against prompt injection
46:45 - Take-away points
49:32 - Questions

[@donatocapitella]

For the moment being, that's absolutely the way to go. I worry about some of the autonomous agents that are becoming all the rage in the media, I love them from a technical point of view and the engineers working at those startups are creative heroes, but we have done fundamental issues still to solve with the underlying tech.

[@micbab-vg2mu]

Many companies are affraid of LLMs and block access and development of any AI tools. What is you opinon about it from the cybersecurity perspective - does not using LLMs prevent as in long run from AI cyberattacks. Of course we do not talking about this creazy idea of autonomus agents - only AI tools for improvment employee productivity as Microsoft Copilot or internal chatbot (without access to the internet) for managing data.

[@donatocapitella]

@@micbab-vg2mu I feel a lot of what upper management in companies are afraid of comes from the hype in the media, that "AI is going to come alive", "human extinction".
If we focus on productivity, LLMs are just another tool, typically integrated in existing tools like Copilot with Office 365 and the issue is more on data confidentiality - but if you're already using Office365 then you've already accepted the risk.
I'd say companies won't be able to fight the use of LLM assistive tools for long.

[@micbab-vg2mu]

Great talk - thank you. I use LMMs every day - but I have never given them autonomy to take control over my systems or data (I promote semi-automation with human in the loop). Human+AI will be more difficult to hack than only AI or only Human.