He Wrote the 27-Year-Old Bug Mythos Found. Then Proved Mythos Wasn’t Necessary | Neils Provos | S2E4
Security Mondays with Eva BennShow Notes
Hi, I'm Eva Benn. Every Monday morning I bring you some of the best global security leaders to talk through what's actually top of mind, the threats they're seeing, the tools they're using, and the decisions they're making right now. Every episode is designed to give you practical things you can go and implement in your organization this week. This week, I sat down with Niels Provos, founder of Security Blueprints and creator of IronCurtain, to discuss how AI is changing vulnerability discovery and what security teams should do about it.The big AI security story this year has been frontier models finding decades old bugs in core software. Niels Provos challenged the idea that this capability is limited to frontier models. Using his open source IronCurtain framework, he reproduced Anthropic’s marquee finding, a 27 year old OpenBSD TCP bug he committed in 1998, with commercial and open weight models.His conclusion: finding vulnerabilities is an orchestration problem, not a frontier model problem.We talked about the growing gap between defenders and attackers, model refusals and rate limits, why working proof of concepts matter for filtering false positives, and how AI could make routine code audits affordable.Niels also made the case for building security invariants, controls that eliminate entire attack classes without requiring human decisions every time. In his analysis of 70 real breaches, three controls would have stopped about 65%.What you’ll walk away with:- Why any model can find the bug should now be the planning assumption- The defender’s tax that attackers on open weight models don’t pay- How to filter AI surfaced findings before they overwhelm your SOC- The real cost of AI code audits- The three controls that could have stopped 65% of 70 analyzed breachesConnect with Niels: LinkedIn: https://www.linkedin.com/in/nielsprovos/ Website: https://www.provos.org/ GitHub: https://github.com/provos Mastodon: https://ioc.exchange/@nielsprovos Resources IronCurtain (open-source): https://github.com/provos/ironcurtain IronCurtain site: https://ironcurtain.dev Finding Zero-Days with Any Model: https://www.provos.org/p/finding-zero-days-with-any-model/ The Day After the Zero-Days: https://www.provos.org/p/day-after-the-zero-days/ "The Day After the Zero-Days" talk (Cloud Security Alliance AI Summit): https://www.youtube.com/watch?v=LKTYZehh_iw Security Blueprints (security invariants): https://securityblueprints.io/ Three invariants / 70-breach analysis: https://securityblueprints.io/posts/three-security-invariants-ciso-challenge/ The Agent Perimeter Fallacy: https://securityblueprints.io/posts/agent-perimeter-fallacy/ Chapters:00:00 — The 27 Year Old Bug AI Found03:43 — Why Mythos Wasn't the Real Breakthrough06:47 — The Workflow That Made AI Vulnerability Discovery Work12:09 — Why AI Agents Can't Be Trusted With Everything25:00 — Why Finding a Vulnerability Isn't Enough29:13 — The Security Controls That Could Stop 65% of Breaches44:08 — The AI Dependency Risk Security Leaders Are MissingWant to hear from more voices in AI security? Security Mondays is a proud partner and supporter of the AI Cyber Magazine. Find it in any Lufthansa or Delta lounges. Not traveling soon? Get your copy here: https://aicybermagazine.comAnd follow along for more free cybersecurity education. 🤗LinkedIn: https://www.linkedin.com/in/evabenn/Instagram: https://www.instagram.com/evabennofficial/TikTok: https://www.tiktok.com/@evabennofficialWebsite: https://www.evabenn.com/
More from Security Mondays with Eva Benn
Remember the headlines about the 3 researchers at Hacktron AI who hacked OpenAI through a photo upload? After $3K in Claude tokens and less than 72 hours they had access to OpenAI’s internal code. Today, we’re hearing from Zayne Zhang, co-founder and CEO of Hacktron AI, and we’re going underneath the headlines into how they did it, how the economics of vulnerability exploitation have changed, and most importantly, what you can do about it.
If you have a hundred findings rated 9.2, which one do you fix first? The score can't tell you. Proof that an attacker can actually exploit it can.
In 2010, Dave DeWalt was CEO of McAfee when one bad update blue-screened computers across 1,672 companies in just 16 minutes. That was one bad update. Now imagine that same kind of accident in the age of autonomous AI agents, systems that can make decisions and take action across production environments at machine speed.
If you watch one video from Season 2 of Security Mondays, make it this one.10 episodes. Some of the most respected names in cybersecurity. Five themes that kept coming up again and again. I boiled it all down to under 15 minutes, so you can walk away with guidance you can use THIS week.Season 2 is officially a wrap. I got to sit down with leaders from completely different backgrounds, with completely different experiences. And they kept landing on the same five lessons. That's not a coincidence.A huge thank you to the incredible guests who made Season 2 possible:Rob Lee: AI security, cybersecurity fundamentals, and how AI is changing the speed of security.Thomas Rocha: AI threat intelligence and how the threats are changing.Helen Oakley & Dimitri Raidman: AI supply chain security and the risks hiding across the AI stack.Neil Provost: Security invariance, and building security that holds up as environments change.Joshua Copeland: Security theater, and why having a control is NOT the same as proving it works.Heath Adams: The attacker mindset, and thinking like the people trying to break your systems.Josh Bressers: The future of open source and the security challenges it faces.Tim Raines: AI aerospace strategy and the security implications of AI in critical environments.Rob Joyce: Offensive operations, national security, and lessons from decades in the field.Every conversation pointed to the same thing:AI is exposing every weak spot in the security playbook we already had, at a speed and scale we can no longer ignore.The five themes:1. AI changed the speed, not the fundamentals.2. Know what you have and what it can access.3. Assume agents will fail and contain the blast radius.4. Stop believing your security works. Test it.5. Finding vulnerabilities is getting cheap. Fixing them is the hard part.But HOW do you actually use any of this on Monday morning?Don't try to do everything at once. Pick ONE lesson. Put it into practice this week.That's how we roll here.Thank you for watching, commenting, and sharing these episodes with your teams. You made Season 2 what it was, and it has meant the world to me. ❤️Season 3 is coming soon, and we're kicking it off with Dave DeWalt on October 5th. Until then, I'll be here every Monday with practical AI security guidance you can turn into action.Which of the five themes hits closest to home for your team? Tell me in the comments 👇If you enjoyed this episode, I'd really appreciate a subscribe and a thumbs up.Follow along for more free cybersecurity education. 🤗---Want to hear from more voices in AI security? Security Mondays is a proud partner and supporter of the AI Cyber Magazine. Find it in any Lufthansa or Delta lounges. Not traveling soon? Get your copy here: https://aicybermagazine.comFollow along for more free cybersecurity education. 🤗LinkedIn: https://www.linkedin.com/in/evabenn/Instagram: https://www.instagram.com/evabennofficial/TikTok: https://www.tiktok.com/@evabennofficialWebsite: https://www.evabenn.com/
Hi, I'm Eva Benn. Every Monday morning I bring you some of the best global security leaders to talk through what's actually top of mind, the threats they're seeing, the tools they're using, and the decisions they're making right now. Every episode is designed to give you practical things you can go and implement in your organization this week.This week, I sat down with Rob Joyce, former Director of Cybersecurity at the National Security Agency and founder of Joyce Cyber. We talked about what changed once AI agents could run most of an attack on their own, and the concrete moves a security leader can make this week to keep pace. What you'll walk away with: - What still stops a top-tier attacker, and the one defense to add now. - What changed in AI-enabled offense, and the assumption to drop from your threat model. - How to check if you qualify for the NSA's free cyber services and sign up this week. - The identity audit to run today: phishing-resistant MFA and killing stale accounts. - The one thing to start Monday, and how to know by Friday that you moved. If you find this helpful, I'd appreciate a thumbs up, subscribe, and share. It helps more people find the show. Connect with Rob: LinkedIn: https://www.linkedin.com/in/rob-joyce-b43445116/ ResourcesUSENIX Enigma 2016 — "Disrupting Nation State Hackers" (talk): https://www.youtube.com/watch?v=bDJb8WOJYdA USENIX Enigma 2016 talk page + transcript: https://www.usenix.org/conference/enigma2016/conference-program/presentation/joyce NSA Cybersecurity Collaboration Center (free services, eligibility): https://www.nsa.gov/ccc Anthropic — "Disrupting the first reported AI-orchestrated cyber espionage campaign" (Nov 2025): https://www.anthropic.com/news/disrupting-AI-espionage Rob Joyce at the Secureframe National Cybersecurity Summit (six moves; AI offense): https://secureframe.com/blog/rob-joyce-dib-cybersecurity Rob Joyce testimony, House Select Committee on the CCP (Mar 5, 2025): https://docs.house.gov/meetings/ZS/ZS00/20250305/117983/HHRG-119-ZS00-Wstate-JoyceR-20250305.pdf XBOW tops HackerOne (2025): https://xbow.com/blog/xbow-on-hackerone-whats-next Chapters:00:00 — Why Hackers Don't Care Who You Are Anymore05:18 — AI Just Changed the Economics of Hacking07:55 — Why Every Organization Is Now a Potential Target10:15 — The Three Defenses You Can't Afford to Ignore17:04 — When AI Agents Become the New Insider Threat20:07 — Why Attackers Can Now Outpace Your Patch Cycle30:01 — The Cryptography Deadline Security Leaders Can't IgnoreWant to hear from more voices in AI security? Security Mondays is a proud partner and supporter of the AI Cyber Magazine. Find it in any Lufthansa or Delta lounges. Not traveling soon? Get your copy here: https://aicybermagazine.comAnd follow along for more free cybersecurity education. 🤗LinkedIn: https://www.linkedin.com/in/evabenn/Instagram: https://www.instagram.com/evabennofficial/TikTok: https://www.tiktok.com/@evabennofficialWebsite: https://www.evabenn.com/