AI Security Podcast

Why do 95% of enterprise AI implementations fail? According to Sandip Wadje, Managing Director at BNP Paribas, many organizations attempt complex reasoning tasks on day one rather than building a mature foundation around data hygiene and simple summarization workflows. In this episode, Ashish sits down with Sandip to explore how global financial institutions navigate Model Risk Management (MRM), GenAI governance, and regulatory expectations across regions like the UK, EU, and US. Sandip breaks down why classical 20-year-old MRM frameworks fall short when applied to non-deterministic black-box LLMs, and why security leaders must focus on output drift and event taxonomies rather than just input prompt filtering. We also examine the concept of the "AI Kitchen" - a cross-functional governance model bringing together IT, CISOs, legal, and Data Protection Officers alongside practical strategies for calculating AI blast radius, cleaning up overprivileged non-human identity (NHI) permissions, and training CSIRT teams for ML SecOps incidents.

Questions asked:
(00:00) Introduction: AI Risk in Regulated Financial Institutions(01:50) Sandip Wadje’s Background at BNP Paribas(02:50) Classical Model Risk Management (MRM) vs. Generative AI(04:40) Governing the Black Box: Finding the Security Delta(08:00) The CMDB Problem: Building an Accurate AI Use Case Inventory(11:30) Why 95% of AI Projects Fail: Summarize, Write, Reason(15:00) Continuous Evaluation (Evals) and Catching Output Drift(18:50) Event Taxonomy: What Happens When AI Decisions Drift?(25:40) Training CSIRT and SOC Teams for ML SecOps Incidents(30:00) Compensating Controls: Remote Browser Isolation & Prompt Monitoring(34:30) Non-Human Identities (NHI) & Cleaning Birthright Permissions(36:30) Balancing a $1M Savings Against a 4% Revenue Fine(38:30) Open-Weight Models vs. Frontier LLMs in Financial Services(41:00) The "AI Kitchen": Cross-Functional AI Governance(44:30) The #1 Rule for AI Security: Understand Your Data First

With over 200 AI security vendors in the market, how does an enterprise CISO decide whether to build a custom solution, buy an off-the-shelf product, or just wait out the hype?
In this episode of the AI Security Podcast, Ashish and Caleb are joined by Kane Narraway, Head of Enterprise Security at Canva, to debate the realities of AI security in modern enterprises. Kane breaks down why simply sandboxing AI agents doesn't work for workforce productivity, explaining that an overly restrictive sandbox renders an agent useless because it inherently needs access to external files and databases to do its job.
We dive deep into the "Confused Deputy" problem, the struggle of granting granular least privilege to AI tools (like letting a bot summarize only Caleb's emails), and whether the old-school concept of network proxies is about to make a massive comeback as the ultimate control layer for AI routing and authorization. Finally, Kane shares why he believes the scariest near-future threat isn't malware, but contractors utilizing "Bring Your Own Agent" (BYOA) in enterprise environments.

Questions asked:
(00:00) Introduction to AI Agents in the Enterprise(01:50) Kane Narraway’s Background (Digital Forensics, Atlassian, Shopify, Canva)(02:50) The Build vs. Buy Debate in the Era of 200+ AI Security Vendors(09:00) Using Wrappers and Harnesses to Control Vendor APIs (Island Browser Example)(11:00) Why GitOps and PRs are Better for AI Configuration than MCP Deployments(13:00) The "Confused Deputy" Problem: Single-Player vs. Multi-Player AI Bots(16:50) How to Handle Agent Identity: "On Behalf Of" (OBO) vs. SPIFFE / NHI(22:50) Why Sandboxing AI Agents Fails for the General Workforce(28:20) Intent-Based Security and the Lack of Granular Access Controls(29:40) Are Proxies the Next Gen Firewall for AI Agents?(34:00) The Terrifying Future of "Bring Your Own Agent" (BYOA)(38:50) The "Gravel Road" Strategy for Managing Shadow IT and Vibe Coding(42:00) Dealing with Vendors Trying to Exploit Shadow IT Land Grabs(49:30) What Security Leaders are Over-Indexing On (Discovery vs. True Access)(50:40) The "You Laugh, You Lose" Cybersecurity Joke Challenge

When AI agents start swarming your enterprise, they won't care about stealth. They will land a beachhead and instantly spawn 500 agents to crawl, probe, and exfiltrate data at machine speed. Is your detection stack ready?
In this episode, Ashish and Caleb sit down with Andy Smith, CEO and co-founder of Tracebit, to completely rethink Deception Technology for the AI era. Forget the heavy, noisy "honeypots" of the 90s. We discuss the modern implementation of deception: lightweight, high-fidelity canary tokens (like fake AWS keys, Chrome cookies, and database tables) that act as guaranteed tripwires the moment an attacker, human or AI, assumes a breach.
Andy shares new research on how you can actively weaponize an AI model's own safety guardrails against it. By embedding specific, controversial text strings (like references to biological warfare or sensitive political events) into decoy secrets.

Questions asked:
(00:00) Introduction to AI Deception(02:30) Andy Smith’s Background and the Founding of Tracebit(03:40) Deception 101: Honeypots vs. Canary Tokens(07:20) The "Assume Breach" Philosophy of Deception(10:00) Why CISOs Default to SIEMs over Quick Deception Wins(13:20) The Psychological Deterrent of Deception on Red Teams(15:10) Setting Up a Database Tripwire (Real-World Example)(17:40) Internal AI Threats: Catching Claude Code in a Production Kubernetes Pod(20:00) Why Deception Fails: The Lack of Strategy and Deployment Complexity(26:30) Using Cloud Serverless (S3/Terraform) to Deploy Deception for Free(28:00) Modern Lateral Movement: Chrome Cookies and Browser History Canaries(41:20) The Future of Attacks: Armies of Fast, Noisy AI Agents(44:50) Weaponizing AI Guardrails to Shut Down Attack Agents(48:20) Where to Start with Your Deception Strategy Today

Resources spoken about during the episode:
- Tracebit Research - Deception warns your teams at the speed of an AI attacker

When the CEO of Anthropic declares that human coding will disappear within six months, followed quickly by the death of software engineering itself, what does that mean for the future of cybersecurity?
In this episode, Ashish and Caleb break down the massive paradigm shift caused by AI coding assistants like Claude Code. Caleb shares his firsthand experience building and deploying software where he has never looked at a single line of the underlying code, arguing that while the need for security will never go away, the humans performing those roles very well might .
We explore the illusion of AI prototyping why building a quick AI tool is easy, but maintaining it in production is a nightmare and dive deep into the "Build vs. Buy" debate . Caleb predicts an upcoming "forest fire" that will wipe out bloated security startups, forcing the market to consolidate around vendors with true, defensible moats based on network effects, hardware integration, or complex regulatory expertise

Questions asked:
(00:00) Introduction(02:50) The Anthropic CEO's Claim: Is Software Engineering Dead? (04:00) Separating Coding from Software Engineering (06:50) Managing Software Without Ever Looking at the Code (08:30) Will AI Eliminate the AppSec Team? (10:30) The Challenge of Legacy Code (COBOL on Mainframes) (15:10) Shifting Focus: From Code Analysis to Agentic Execution (18:00) The Coming "Forest Fire" in the Security Startup Landscape (21:00) The "Build vs. Buy" Illusion: Prototyping vs. Production (36:30) How to Build a Defensible Moat in AI Security (41:00) Why Hardware and Red Tape Are the Ultimate Moats (46:30) The AI Scaffolding Approach for Enterprises (47:50) Automating SIEM Detections

Resources spoken about during the episode:
World Economic Form - Davos 2026

Why do CISOs still struggle with asset intelligence in 2026? Despite decades of security tooling, most organizations still have a massive 40% "dark matter" blind spot in their environment and the explosion of ephemeral AI agents is only making it worse.
In this episode, Ashish and Caleb sit down with Joe Diamond, CEO, Axonius to discuss the evolution of the asset space. We explore why traditional CMDBs (which track business processes and IT hardware) fall short for cyber asset attack surface management (CAASM), and why the industry is shifting from static asset inventory to dynamic asset intelligence. Joe spoke about how AI agents whether they run for five minutes or five months must be treated as a distinct asset class, complete with their own access logs and token utilization tracking.
The conversation also goes into the future of enterprise software interfaces. Joe predicts that within three to five years, the traditional dashboard UI will completely disappear, replaced entirely by natural language prompts and AI-driven BI. Finally, we tackle the "Build vs. Buy" dilemma: if AI can integrate tools in five minutes, why do we still need vendors?

Questions asked:
(00:00) Introduction(01:50) Joe Diamond's Background and Journey into Cybersecurity(02:50) Why Asset Management is Still an Unsolved Problem(04:00) The 40% "Dark Matter" Blind Spot in Enterprise Environments(05:30) How Do We Actually Define an Asset?(08:30) CMDB vs. Asset Intelligence: Understanding the Delta(12:30) Defining AI Models and AI Agents as an Asset Class(15:30) Do Ephemeral AI Agents Need to be Tracked?(18:30) The "Time Machine" Feature: Tracking Asset Configuration Drift(20:30) Use Case: Remediating the CrowdStrike Outage Using Asset Intelligence(23:30) Why You Need Asset Intelligence if You Already Have CSPM/CNAPP(31:30) The End of the UI: Why Dashboards Will Be Replaced by AI Prompts(36:30) A Simple 3-Question Framework for AI Asset Management(38:30) Build vs. Buy: Why AI Cannot Operate and Maintain Software