

EP198 GenAI Security: Unseen Attack Surfaces & AI Pentesting Lessons
11 snips Nov 11, 2024
Ante Gojsalic, Co-Founder & CTO at SplxAI, dives into the intricacies of securing generative AI applications. He outlines the unique challenges of penetration testing in this realm, such as non-determinism and the complex interplay of data and applications. Ante discusses the most concerning current attack surfaces and shares his insights on common security mistakes companies make. He emphasizes the importance of blending automated pentesting with human expertise and offers practical strategies for learning about AI security. Tune in for crucial tips on navigating this evolving landscape!
AI Snips
Chapters
Transcript
Episode notes
GenAI Pentesting Challenges
- Pentesting Generative AI apps is non-deterministic, requiring multiple tests with varied inputs.
- Unlike traditional apps, Generative AI apps embed logic and data within models, making role-based access control difficult.
Data in Generative AI Apps
- Data is intertwined with Generative AI apps, unlike traditional apps where data can be separated for testing.
- This makes it impossible to test Generative AI apps without real or representative data.
Multimodality Complications
- Multimodality in Generative AI apps means the same input can yield different results depending on the modality (text, image, voice).
- A jailbreak might work via image input but not text input on the same chatbot.