SOBRE ESTE EPISÓDIO
What happens when an AI model can find more vulnerabilities in a day than a red team could find in a year?
Welcome to Razorwire, the podcast where we share our take on the world of cybersecurity with direct, practical advice for professionals and business owners alike. I'm Jim and in this episode, I'm joined by Martin Voelk, penetration tester and AI red teamer, and Jonathan Care, lead analyst at KuppingerCole covering AI and cybersecurity.
Anthropic recently announced Mythos, a security-focused AI model reportedly capable of discovering vulnerabilities that have gone undetected for decades, including a 27-year-old bug in OpenBSD. But how much of this is genuine breakthrough and how much is marketing? This episode cuts through the hype and asks what Mythos actually means for the cybersecurity industry, from the arms race it signals between AI model providers to the competitive implications of restricting access to a small group of US-based companies.
The conversation goes well beyond Mythos itself, into the reality that AI-powered hacking at scale is already happening, that existing models have already been used to compromise government infrastructure, and that open source and non-Western alternatives are freely available to anyone who wants them. With 80% of code now being vibe coded with minimal security checks, jailbreaking tools available on the open web and CISOs unable to keep pace with the speed of adoption, the question isn't whether AI will change cybersecurity. It's whether the industry can adapt fast enough to survive what's already here.
Three key talking points:
- The Mythos hype vs the reality of AI-powered hacking: Anthropic's announcement made headlines, but the capability to find and exploit vulnerabilities at scale already exists in models available to anyone. This episode asks whether Mythos is really the breakthrough it's been presented as, or whether the industry should be more concerned about what's already out there, including a recent attack on the Mexican government carried out entirely using standard AI models.
- The competitive and geopolitical implications of restricted AI models: Mythos has been restricted to a small group of US-based companies, giving at least one major EDR vendor a significant edge over every competitor. But by announcing the capability publicly, Anthropic has effectively told the rest of the world it's possible to build. With Chinese, Russian and open source models already filling the gap, the question is whether restricting access to Western models actually contains anything at all.
- Why security practitioners can't keep up and what comes next: The pace of AI development has outstripped the ability of security teams to keep up. Even full-time practitioners can't stay on top of the daily volume of new models, new vulnerabilities and new attack techniques. If the people doing this for a living are struggling, what chance does an SMB with a part-time security person have? And where does it end? Possibly with offensive and defensive AI agents fighting it out at scale, with humans increasingly on the sidelines.
Whether Mythos lives up to the hype or not, the arms race it signals is already underway. If you want to understand what that means for cybersecurity, this is the conversation to listen to.
On the implications of restricting AI security models:
“Anthropic may be doing this, but for those of us who are not lucky enough to be Anthropic's friend, other countries, other organisations are not so circumspect.”
Jonathan Care
Listen to this episode on your favourite podcasting platform: https://razorwire.captivate.fm/listen
In this episode, we covered the following topics:- Anthropic's Mythos Announcement Find out what Anthropic is claiming about Mythos, why it reportedly found a 27-year-old vulnerability in OpenBSD and why not everyone is convinced it's the breakthrough the headlines suggest.
- AI-Powered Vulnerability Discovery at Scale Understand why the ability to find and exploit vulnerabilities at machine speed already exists and why Mythos may be less of a leap forward than it first appears.
- The Mexican Government Hack Hear how standard, publicly available AI models were used to compromise multiple government entities and exfiltrate massive amounts of sensitive data over a matter of weeks, without any zero days involved.
- Restricted Access and Competitive Advantage Explore why limiting Mythos to a handful of US-based companies raises questions about competitive fairness and what it means when one EDR vendor gets capabilities that nobody else has access to.
- The Open Source and Non-Western Model Landscape Discover why restricting Western models may not contain much at all, with Chinese, Russian and uncensored open source alternatives already being used by security researchers and attackers worldwide.
- Vibe Coding and Unchecked AI-Generated Code Find out why an estimated 80% of code is now vibe coded, why most of it isn't being properly tested and what that means for the attack surface organisations are unknowingly building.
- Jailbreaking and Uncensored Models Learn why tools that can jailbreak frontier models on the fly are freely available on the open web and what that tells us about the limits of trying to restrict AI capability.
- The CISO's Impossible Position Understand why CISOs are caught between an industry that's moving faster than they can govern and organisations that want to adopt AI regardless of whether the security is ready.
- Keeping Up With the Pace of Change Explore why even full-time security practitioners are struggling to stay on top of the daily volume of new developments and what that means for organisations with fewer resources.
- The Future: Agent vs Agent Hear why the near future of cybersecurity may look less like humans defending networks and more like offensive and defensive AI agents battling it out at scale, with practitioners increasingly in a supervisory role.
Resources Mentioned
Anthropic – Mythos/Project Glasswing
Mexican Government Cyberattack
GodMode AI / Pliny the Prompter (jailbreaking harness)
Hugging Face (uncensored models)
SpartanX Technologies / SpartanX AI
Connect with your host James ReesHello, I am James Rees, the host of the Razorwire podcast. This podcast brings you insights from leading cyber security professionals who dedicate their careers to making a hacker’s life that much more difficult.
Our guests bring you experience and expertise from a range of disciplines and from different career stages. We give you various viewpoints for improving your cyber security – from seasoned professionals with years of experience, triumphs and lessons learned under their belt, to those in relatively early stages of their careers offering fresh eyes and new insights.
With new episodes every other Wednesday, Razorwire is a podcast for cyber security enthusiasts and professionals providing insights, news and fresh ideas on protecting your organisation from hackers.
For more information about us or if you have any questions you would like us to discuss email podcast@razorthorn.com.
If you need consultation, visit www.razorthorn.com, We give our clients a personalised, integrated approach to information security, driven by our belief in quality and discretion.
LinkedIn: Razorthorn Security
YouTube: Razorthorn Security
TikTok: Razorwire Podcast
Instagram: Razorwire Podcast
Twitter: @RazorThornLTD
All rights reserved. © Razorthorn Security LTD 2025
