As AI systems become more integrated into enterprise operations, understanding how to test their security effectively is paramount.

In this episode, we're joined by Leonard Tang, Co-founder and CEO of Haize Labs, to explore how AI red teaming is changing.

Leonard discusses the fundamental shifts in red teaming methodologies brought about by AI, common vulnerabilities he's observing in enterprise AI applications, and the emerging risks associated with multimodal AI (like voice and image processing systems). We delve into the intricacies of achieving precise output control for crafting sophisticated AI exploits, the challenges enterprises face in ensuring AI safety and reliability, and practical mitigation strategies they can implement.

Leonard shares his perspective on the future of AI red teaming, including the critical skills cybersecurity professionals will need to develop, the potential for fingerprinting AI models, and the ongoing discussion around protocols like MCP.

Questions asked:

• 00:00 Intro: AI Red Teaming's Evolution

• 01:50 Leonard Tang: Haize Labs & AI Expertise
• 05:06 AI vs. Traditional Red Teaming (Enterprise View)
• 06:18 AI Quality Assurance: The Haize Labs Perspective
• 08:50 AI Red Teaming: Real-World Application Examples
• 10:43 Major AI Risk: Multimodal Vulnerabilities Explained
• 11:50 AI Exploit Example: Voice Injections via Background Noise
• 15:41 AI Vulnerabilities & Early XSS: A Cybersecurity Analogy
• 20:10 Expert AI Hacking: Precisely Controlling AI Output for Exploits
• 21:45 The AI Fingerprinting Challenge: Identifying Chained Models
• 25:48 Fingerprinting LLMs: The Reality & Detection Difficulty
• 29:50 Top Enterprise AI Security Concerns: Reputation & Policy
• 34:08 Enterprise AI: Model Choices (Frontier Labs vs. Open Source)
• 34:55 Future of LLMs: Specialized Models & "Hot Swap" AI
• 37:43 MCP for AI: Enterprise Ready or Still Too Early?
• 44:50 AI Security: Mitigation with Precise Input/Output Classifiers
• 49:50 Future Skills for AI Red Teamers: Discrete Optimization

Resources discussed during the episode:

Baselines for Watermarking Large Language Models

Haize Labs