Cheese Rolling and Frontier AI

When I started putting this article together, I was trying to find something funny, something personal, something non-AI, that would serve as an introduction to a serious topic. So what better topic to link the challenges of near-frontier AI, cybersecurity, and everything that comes with it, than the good old British sport of cheese rolling.

Stick with me.

Cheese rolling is a sport where people climb to the top of Cooper's Hill, a wheel of Gloucestershire's finest gets released, and a bunch of random nutters from around the world start running down an extremely steep hill trying to catch it.

What's the link to the serious topic? Well, now that the AI genie's been let out of the bottle, we, and I'm talking about technology professionals here, are all desperately trying to keep up and catch the cheese. The only difference is the hill seems to keep getting steeper, and the cheese keeps moving faster.

We've run a handful of Frontier AI Cyber Readiness Assessments this year. Here's what chasing the cheese actually looks like once you're on the ground.

Build app security into development now, not later

App development is moving fast. CI/CD, multi-cloud, agentic development are all standard now, but application security still hasn't landed as a real "thing" in sprints and the backlog. If developers aren't won over, it doesn't happen. Not the security review, not the cloud configuration checks, not the latest images.

This isn't a nice-to-have anymore. Agentic development is shipping more code, faster, because the business demands it, everyone's releasing quickly so you have to be in it to win it. No fault or blame, that's just compressed timelines and raised expectations doing what they do. But it means more vulnerabilities going in the door than any team can review by hand. Appsec has to be embedded at the point code gets written, not bolted on before release. Shift left, properly, not as a slogan on a slide.

Get on top of non-human identities and agent permissions, urgently

This is the one most organisations are furthest behind on. Every agent, every service account, every automated pipeline is a non-human identity, and most of them were provisioned with whatever access got the project moving, not what the task actually required.

That matters because of how these attacks actually play out. A model finds a vulnerability, chains it to another one, then another, then finds an overprivileged non-human identity, and that's the escalation. If Claude Mythos and OpenAI's own models can escape sandboxes with all the investment, guardrails, and security those labs have built in, ordinary organisations need to assume their own environment offers a lot less resistance.

Inventory your non-human identities properly. Know what every agent and service account can actually touch, not what you assumed it could touch. Scope permissions to the task, not the project. This is urgent, not roadmap material.

Prioritise your vulnerability backlog by chainability, not just severity

Near-frontier open models are closing the gap to closed frontier models fast, four to seven months now. Some run on laptops or lower-spec hardware. That means the barrier to entry for serious offensive capability isn't high anymore, and it's getting lower. Cybercriminals can download these models, strip the guardrails, fine-tune them, sell them on the dark web. Jailbroken models pointed at your endpoints don't get tired, andneverget bored.

This changes what your backlog actually means. Your medium and low vulnerabilities, the ones every team quietly deprioritises, are exactly what these models are good at chaining together into something serious. Stop triaging purely on CVSS score. Start asking what combines with what, and close the gaps that create a path, not just the ones with the scariest individual rating.

Close the gap between your SOC's speed and the attack's speed

Most SOCs are still running on tools that were genuinely good three or four years ago, SIEM, some SOAR, 24x7 coverage, escalation processes that assume a human is in the loop making judgement calls. All sound practice. None of it moves at the speed a machine-speed attack does. Boom, done, while your escalation process is still routing the alert.

You don't need to throw out what you've built. You need to find where the human bottleneck actually sits, usually triage and initial escalation, and put automation or AI-assisted response there specifically, rather than trying to modernise everything at once.

Extend the assessment to your vendors, not just yourself

Zoom out and the same loop is running inside every organisation you depend on. Same development speed pressure, same appsec gap, same SOC lag. The products and platforms in your stack are carrying the same exposure you are, and you inherit it the moment you take a dependency, an integration, or an update from them. Don't stop your review at your own boundary. Ask the same questions of your critical vendors that you're asking of yourself.

Fund this as an ongoing concern, not as a fixed roadmap

Good cybersecurity fundamentals, network segmentation, knowing your assets, identities and data, knowing your critical systems, risk-prioritised patching, are all still valid. They just have to happen faster, and the funding model has to reflect that. A 24-month roadmap built around today's threat model will be out of date in twelve, because the models driving this keep shipping. Build the business case with an actual quantitative view of the cost of these risks, and check it against your cyber and operational resilience insurance gaps. That's what gets budget released for something ongoing rather than a one-off project.

Be flexible. Be nimble. Be open to course correct. Be wary of sunk cost fallacies.

The cheese doesn't wait for you to tie your shoelaces.