The AI Security Podcast

This week I discuss a lecture by Jack Clark, check it out here. And check out his newsletter Import AI. https://www.youtube.com/watch?v=8zIcP5WlShw

This week we're finally delving into Claude Mythos and Project Glasswing! We discuss both the technical and geostrategic challenges, and what it means for the future of AI security.

The announcement: https://red.anthropic.com/2026/mythos-preview/

Glasswing: https://www.anthropic.com/glasswing

AISI: https://www.aisi.gov.uk/blog/our-evaluation-of-claude-mythos-previews-cyber-capabilities

And if you're interested in my book: https://nostarch.com/practical-ai-security

AI Security Fundamentals Course: https://aisecurityfundamentals.com/

If you’re trying to break into AI security, it can feel confusing — do you need to be a machine learning expert, a cybersecurity professional, or both? In this episode, we break down practical tips for getting hired in AI security, from the skills that actually matter to the types of projects and experience that can help you stand out. We discuss how to build relevant expertise in areas like adversarial machine learning, AI risk, and model security, as well as how to position yourself for roles in startups, research labs, and large tech companies. Whether you’re coming from a cybersecurity, data science, or general tech background, this episode will give you actionable advice on how to start building a career in one of the fastest-growing areas of technology. 🚀

Want to give a great conference talk (and not bore everyone to death)? In this episode, I share practical tips for giving a strong conference talk — from structuring your idea to actually delivering it on stage. #PublicSpeaking #Conferences #CFP #TechTalks #Cybersecurity #AI

In this episode, we dig into model weight security — what it means, why it’s emerging as a critical issue in AI security, and whether the framing in the recent RAND report on securing AI model weights actually helps defenders and policymakers.

We discuss the RAND report Securing AI Model Weights: Preventing Theft and Misuse of Frontier Models — exploring its core findings, including how model weights (the learnable parameters that encode what a model “knows”) are becoming high-value targets and the kinds of attack vectors that threat actors might use to steal or misuse them.

#ai #aisecurity #cybersecurity 👉

Read the full RAND report here:https://www.rand.org/pubs/research_reports/RRA2849-1.html

In this episode, we dig into Model Context Protocol (MCP) and agent-to-agent (A2A) communication — what they are, why they matter, and where the real risks start to emerge.We cover:- What MCP actually enables beyond “tool calling”- How A2A changes the threat model for AI systems- Where trust boundaries break down when agents talk to each other- Why existing security assumptions don’t hold in agentic systems- What practitioners should be thinking about now (before this ships everywhere)This one’s for anyone working on AI systems, security, or governance who wants to understand what’s coming before it becomes a headline incident.As always: curious to hear your takes — especially where you think the biggest risks (or overblown fears) really are.

As promised, I’m back with Tania for a deep dive into the wild world of agentic AI security — how modern AI agents break, misbehave, or get exploited, and what real case studies are teaching us.

We’re unpacking insights from the Taxonomy of Failure Modes in Agentic AI Systems, the core paper behind today’s discussion, and exploring what these failures look like in practice.

We also break down three great resources shaping the conversation right now:

Microsoft’s Taxonomy of Failure Modes in Agentic AI Systems — a super clear breakdown of how agent failures emerge across planning, decision-making, and action loops: https://cdn-dynmedia-1.microsoft.com/is/content/microsoftcorp/microsoft/final/en-us/microsoft-brand/documents/Taxonomy-of-Failure-Mode-in-Agentic-AI-Systems-Whitepaper.pdf

OWASP’s Agentic AI Threats & Mitigations — a practical, security-team-friendly guide to common attack paths and how to defend against them: https://genai.owasp.org/resource/agentic-ai-threats-and-mitigations/

Unit 42’s Agentic AI Threats report — real-world examples of adversarial prompting, privilege escalation, and chain-of-trust issues showing up in deployed systems: https://unit42.paloaltonetworks.com/agentic-ai-threats/

Join us as we translate the research, sift through what’s real vs. hype, and talk about what teams should be preparing for next 🚨🛡️.

A quick end-of-year message to say thanks. Thanks for being part of the channel this year — whether you’ve been watching quietly, sharing, or arguing with me in the comments. I really appreciate it.I hope you have a good Christmas and holiday period, whatever that looks like for you. Take a break if you can. See you in 2026.

In this episode, I’m joined by Bandana Kaur — a cybersecurity researcher, speaker, and all-round superstar who somehow managed to do in her teens what most people are still figuring out in their thirties. 🤔

Bandana is just 18 years old, entirely self-taught in cybersecurity, already working in the field, and recently gave three talks at Black Hat. Yes, three! 😱

We talk about how she taught herself cybersecurity as a teenager, how she broke into the industry without a traditional pathway, and what it’s actually like being young (and very competent) in a field that still struggles with gatekeeping. Bandana shares what she focused on while learning, how she approached opportunities like conference speaking, and what she thinks matters most for people trying to get into security today.

This conversation is part career advice, and part reminder that you don’t need permission — or a perfectly linear path — to do meaningful work in cybersecurity.

Follow Bandana: @hackwithher

Remember that time I invited myself over to Greg's place with my camera? This is part 2 from that great conversation. I'm curious to hear whether you've heard a lot about EA? It's something really big in the AI world but I'm conscious a lot of people outside the bubble haven't heard of it. Let me know in the comments! Check out Greg's work here: https://www.goodancestors.org.au/

MIT AI Risk Repository: https://airisk.mit.edu/

The Life You Can Save (book): https://www.thelifeyoucansave.org/book/

80,000 hours: https://80000hours.org/

Learn more about AI capability and impacts: https://bluedot.org/