1 of 9

Safety in Multimodal AI - Panel

Klaudia Kloc

CEO Vidoc Security Lab

Samantha Wylatowska

Sandra Mau

Solutions Architect

VP Cloud & IP Government Solution

Narcisse Zekpa

Sr. Solutions Architect

2 of 9

Key Security Risks introduced by Multimodal Large Language Models (MLLM)

Adversarial Attacks

Data Leakage

DeepFakes

Subtle input manipulations to trick AI into errors and/or unsafe output

Diversity in data types increase privacy risks

Sophisticated and media rich content undermines authenticity

Background and Anecdotes

Could you please introduce yourself and share a bit about your professional background? Please share a specific anecdote where multimodal AI posed safety challenges?
What measures or techniques did you implement? How effective were they?

Managing multiple media types

Vidoc Security already detects code vulnerabilities especially within LLM generated code, with the number of modalities growing, do you see additional challenges specific to certain content types (images vs video vs audio)?
Do you foresee threats spanning beyond code into other content types (diagrams, images, .md files …)

Securing Agents

With the explosion of Agents can you share some insight on specific challenges related to multimodal data assets access management and governance?
What new challenges (if any) does multi-modality bring to the table?

Copyright and IP Protection

Terms & Conditions of well-known models are relatively loose and push back the copyright protection responsibility onto model users. Can you share your insight along with some best practices firms should be aware of to protect themselves in that regard?

Regulation and Governance

How can we design AI regulations that remain effective and relevant as the technology rapidly evolves, while avoiding rules that might stifle beneficial innovation or become quickly outdated ?

Which regulatory standards do you see being already implemented in your industry?

Parting thoughts and looking ahead

Parting thoughts or best practices/tips you would like our audience to leave with?
Fast forward 10 years from now what do you envision?

Audience Q & A