Show Notes
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
More from AI Security Podcast
While attackers using AI can now operate at "machine speed," traditional enterprise security teams are often stuck moving at "meeting speed." To bridge this gap, organizations must fundamentally rethink how they structure their security operations, leverage AI for defensive scale, and eliminate the bureaucratic bottlenecks that delay critical patching and remediation. In this episode of the AI Security Podcast, Ashish and Caleb sit down to analyze the September 2026 Anthropic Threat Report and the underlying causes of recent high-profile AI sandbox escapes. Caleb explains how security leaders can fight AI with AI, sharing how he successfully used autonomous agents to conduct a 98% accurate least-privilege architectural audit across his entire organization. Caleb also speaks about the "barbell hiring" strategy eliminating middle management to build a highly agile, cross-functional Security SWAT team composed entirely of principal-level experts and junior engineers.
As the speed of automated threats outpaces traditional patching cycles, organizations must prioritize structural containment over purely reactive detection. Cloud network architectures frequently default to permissive egress, meaning workloads can reach the public internet unless explicitly restricted. In this episode, Doug Merritt, CEO of Aviatrix and former CEO of Splunk, details how to build practical containment strategies for modern cloud and AI environments. He outlines how security teams can implement egress controls in monitor mode to validate traffic without risking system downtime. The discussion covers the transition from legacy IP-based firewalls to identity-based policies for ephemeral serverless and Kubernetes workloads. By prioritizing reachability and blast radius as core metrics, organizations can systematically reduce their attack surface and improve resilience.
When non-technical employees can spin up fully functioning apps in days and AI agents are running parallel workflows that babysit their own pull requests, the traditional SDLC is dead. In this episode, Ashish sits down with Prahathess Rengasamy, Senior Security Engineer at SpaceX AI, to discuss how the staggering compression of development time requires an entirely new approach to security engineering. Prahathess explains why legacy 10-minute PR wait times are incompatible with multi-agent workflows, and how security teams can build internal Model Context Protocol (MCP) servers to inject compliance and security guardrails directly into an agent's context window. We unpack the differences between Q&A chatbots and autonomous coding agents, the critical need to preserve foundational preventative cloud controls (like AWS SCPs and GCP Org Policies), and why holding an autonomous machine accountable for management decisions is a critical architectural flaw. Finally, Prahathess breaks down the "build vs. buy" debate, revealing how small security teams can leverage AI to rapidly deploy internal platforms and paved roads that previously took months to engineer.
How does a global financial engine moving trillions annually manage cyber risk in the age of AI? In this episode, Ashish sits down with Subra Kumaraswamy, CISO of Visa. Subra discusses Visa's journey as part of Anthropic's Glasswing Project testing the Mythos model, revealing why a majority of Mythos-discovered vulnerabilities were non-exploitable due to robust zero-trust architecture and micro-segmentation. Subra explains why Visa open-sourced VVAH (Visa Vulnerability Agentic Harness) to help security teams scale vulnerability discovery, prioritization, and remediation across any model. He also breaks down how Visa triages 98% of Level 1 SOC incidents with AI agents, why "Mean Time to Adapt" (MTTA) is replacing legacy patch timelines, and how cross-functional AI governance enables innovation while maintaining strict production controls
Is the AI "vulnpocalypse" already here? According to Casey Ellis, Founder of Bugcrowd and pioneer of Disclose.io, we aren't quite in an apocalypse yet, we're actually in a "slopdemic." The cost of discovering vulnerabilities has plummeted, flooding bug bounty and SOC triage queues with low-quality, noisy submissions. Because these queues are so overwhelmed, many highly skilled researchers are simply hoarding zero-days because reporting them has become too difficult. In this episode, Ashish sits down with Casey to unpack the major themes and mindset shifts from RSA and Black Hat 2026. Casey breaks down how AI is shrinking the OODA loop for defenders, forcing the industry to adopt a true "assume breach" mentality and reconsider deception technology to frustrate active adversaries. They also explore the risks of non-technical employees "vibe coding" corporate applications and why CISOs must get hands-on with AI tools at home if they want to understand the risks their workforce is taking.