AI Thinking

Your AI Strategy has a 27-year-old bug in it.

You just don’t know yet.

Katie Lips, CEO and Founder of Rebel Futures/17 May 2026
Your AI Strategy has a 27-year-old bug in it.

Anthropic’s Mythos isn’t just a cybersecurity story or just an AI story, or just a marketing strategy story… it’s a business story, that incorporates leadership, strategy, security, engineering, product and AI together.

Major cyber events of the last decade followed the same pattern of boardroom alarm, followed by a regulatory response and some updated compliance frameworks. Target tightened procurement rules, Equifax accelerated GDPR enforcement, and SolarWinds produced a US executive order and made supply chain security briefly fashionable. Each one generated the right kind of news, but none of them fundamentally changed how organisations connect cybersecurity to business strategy. Cybersecurity made the headlines, had its 15 minutes of fame, and then disappeared back into the shadows.

With Mythos, though, we need to take things (cybersecurity things) more seriously.

About a decade ago, I started working with a client who wanted to build a consumer cybersecurity product. It was an unusual brief at the time, and it opened a door I’ve never fully closed. What struck me then, and has stayed with me through many engagements since, was how fundamentally misunderstood cybersecurity was as a business problem. It is treated as a technical specialism, and very often siloed away from pretty much everything else. I can see why; it’s difficult, specialised, and sometimes even depressing, especially when you learn, as a non-technical person just how many threat vectors there are in the systems and tools we use every day and just how many bad actors there are desperate to exploit them.

Yet the insight driving my client’s consumer product was simple and right: everyone has a role to play in cybersecurity: individuals, teams and organisations; and from UX and design through product to IT and to the siloed cybersecurity team. It was a visionary product in many ways, as it brought security thinking out of the shadows.

Since then, I’ve worked in other organisations that align engineering with security, and even in some that conceptually connect Product and UX to security. In my time at Which?, I encouraged businesses to build customer data protections in at the UX level up. But it’s true to say most organisations don’t think in this way and cybersecurity is confined to its own special ‘dark arts’ corridor.

Now Anthropic has released, or more accurately, carefully not released, a model called Mythos, and every structural weakness that’s been quietly accumulating in how businesses think about cybersecurity has just been exposed. Cybersecurity, or rather a lack thereof has been in the news. Is every system we use now unsafe because of big, bad AI? Should we all be worried, from boardroom execs to end users?

What Mythos does

The hype has gotten ahead of the facts in some corners, so let’s be precise about what Mythos does and doesn’t do.

Claude tells me “Mythos Preview is the current version of the model; Anthropic’s most capable model to date; a general-purpose LLM whose cybersecurity abilities emerged as a byproduct of improvements in coding and reasoning, not as an intentional design goal.”

Of course, it turned out that finding and exploiting software vulnerabilities is exactly the kind of multi-step, pattern-matching, problem-solving that powerful reasoning engines are good at.

During testing, “Mythos turned vulnerabilities found in Firefox’s JavaScript engine into working exploits 181 times out of several hundred attempts”. Whereas its predecessor managed it twice. It found zero-days across every major operating system and every major web browser. It found bugs that were 10, 15, 20 years old hidden in plain sight in codebases that thousands of engineers had worked on and audited for decades.

A 27-year-old vulnerability in OpenBSD, one of the most security-hardened operating systems ever built, was found and weaponised overnight - not by a nation-state hacking team, but by a language model. The step change is that engineers with no formal security training were able to direct it to find vulnerabilities overnight.

The attackers just got a thousand times faster. Your defence didn’t.

So what is Mythos’ impact? To understand the scale of the innovation, you have to understand a little about previous threat scales. The cybersecurity industry has always operated on the implicit assumption that serious offensive capability requires serious expertise. Nation-states, elite hacking crews, well-funded criminal organisations were the only actors capable of finding and weaponising novel vulnerabilities at scale. Have-a-go hackers weren’t really a threat.

Mythos breaks that assumption. Now, instead of decades of experience, a bad actor would just need access to a model like Mythos. This is not a gradual shift but a huge issue for organisations that were already underinvested in cyber, which, let’s face it, is most of them.

Two problems colliding

In parallel to the decade I’ve spent watching cyber stay frustratingly siloed from the rest of business, I’ve watched AI strategy make a different mistake: without a strategic approach, it appears everywhere, and all at once, but only at the surfaces.

The low-hanging fruit of enterprise AI right now is operational efficiency; automating workflows, cutting costs, improving product UX. These are real wins and I’m not dismissing them. But they represent AI applied to the visible, it’s AI at the edges. Perhaps even performative.What not enough organisations are doing is applying AI across the full stack, including the unglamorous layers where there’s structural risk, e.g. in the legacy systems. Everyone has legacy. But using AI solely to innovate at the fancy end of product while leaving the rest of the stack unexamined isn’t just a missed opportunity; in a post-Mythos world, it’s a liability.

Many businesses already have AI embedded in their security tools: CrowdStrike, Darktrace, Microsoft Sentinel and others have been using AI for threat detection and anomaly spotting for years. But having AI in your security tools is not the same as having AI in your security strategy.

AI in the platforms used my most InfoSec teams is largely reactive; it detects threats that have already materialised. What Mythos represents is something categorically different: proactive, deep, code-level vulnerability research that finds weaknesses before anyone exploits them. It’s the difference between a smoke alarm and a structural audit of the building. Most businesses have the smoke alarm, whereas very few have done the audit.

For many organisations, cyber never got properly connected to software engineering. For most, AI strategy still hasn’t reached the security layer.

Mythos doesn’t just alter your threat landscape, and as a result your cybersecurity strategy. It changes what a complete AI strategy looks like. Any organisation with an AI roadmap that doesn’t include a defensive cyber component is now, by definition, operating with an incomplete plan. Before Mythos, you could build an AI strategy that didn’t mention cybersecurity and be broadly defensible. That’s no longer true.

Connecting cyber to business strategy doesn’t need to mean your CEO has to become a security expert. For me, it means three things:

  1. Security has a seat in product and engineering decisions from the start, not after the fact;
  2. AI is deployed defensively across the stack, not just at the product surface; and
  3. Cybersecurity teams engage in conversations about business risk, not just technical compliance.

What this means if you’re running a business

If you’re in the C-suite, the immediate question is liability and tempo. Patch cycles (to fix vulnerabilities you find) historically take months, whereas now Mythos-class models find and weaponise vulnerabilities in hours. The gap, between the speed of discovery and the speed of remediation, is now your strategic problem, not your IT team’s operational one.

Modern SaaS products are built on layers of open-source dependencies including libraries, frameworks, and code written years or decades ago by teams with limited security budget. Mythos scanning an OSS repository is like setting a genius loose on a problem nobody has properly looked at in fifteen years. There will be issues in your stack. That’s not speculation… it’s probability.

The businesses that will lead in this environment are those that finally break down the siloes; that stop treating cyber as the IT team’s problem and start treating it as a product problem, an engineering problem, a commercial problem. That understand AI not as a tool for making the front end shinier, but as infrastructure that needs to run all the way down: through the stack, through the organisation, through the layers that have been quietly accumulating risk for years.

Mythos changes how you build from here

This is the part most strategic conversations about Mythos are missing. It doesn’t just change how you think about what you’ve already built. It (powerful AI) changes how you should build from this point forward.

If AI can scan your codebase and surface exploitable vulnerabilities faster than your team can ship new features, then security can no longer be a phase that comes after development. It has to be the same conversation: embedded in every architectural decision, every dependency choice, every code review. The calculation around which open-source libraries you adopt, how quickly you can deploy patches, and how deeply AI is integrated into your development pipeline has fundamentally changed. It impacts engineering teams, it impacts Product teams. “AI thinking” (which I will write more about soon), needs to be embedded across.

It doesn’t matter that Mythos is currently held safely in the hands of a chosen few under Project Glasswing. Mythos is the first, not the only, not the last. The organisations that adapt their engineering culture to this reality now, will have a structural advantage over those that treat cyber as a compliance retrofit later.

Glasswing is indeed billed as ‘Securing software for the AI era’ - assuming there will always be software that has not been built by AI. There will always be software that needs securing. And as software becomes probabilistic (and changes), it will need securing and re-securing as it learns from itself.

This isn’t a “monitor the situation” moment.

Mythos does not just represent a threat. Rather, it’s a competitive signal. The companies that come out of this period strongest won’t be those that respond reactively when an incident forces their hand. They’ll be the ones that treat this as a call to action - to finally connect cyber to strategy, to deploy AI where it really matters (through the stack), and to build a security posture that becomes a genuine differentiator.

The question to ask isn’t whether your business has vulnerabilities: because, of course it does. It’s whether you have the leadership to find them before someone else does.

Some leaders will read this, forward it to the ‘right’ person, and move on. A few will recognise it for what it is: the moment the old model of treating cyber as an independent silo ended. The gap between those two groups is about to become very visible, very fast.

Katie Lips is CEO and Founder of Rebel Futures. She helps organisations grasp the opportunities at technology inflection points, from the early web and pre-iPhone mobile to cloud-based SaaS and now AI. She has held senior roles at EY and Aimia, and was Head of Digital Policy at Which?. Katie holds a Harvard Kennedy School qualification in Leading in Artificial Intelligence and is a Founder Institute graduate.

Related: explore our work on AI Strategy.

base44
Edit with Base44