Maksim Kabakou - Fotolia
Stop debating frontier AI – start defending against it
The Computer Weekly Security Think Tank considers if Anthropic’s Claude Mythos frontier AI model is a benefit or barrier to achieving resilient enterprise IT security, and how security leaders need to adapt.
In mid-June, global access to two of the most anticipated frontier AI models was abruptly suspended following a US Commerce Department export control directive, a development few in the industry had seen coming.
This has prompted significant discussion across the technology industry, with implications that extend well beyond any single vendor. However, as a security leader, my instinct is always to look past the immediate story. Right now, the bigger picture demands our attention.
In this case, the bigger picture shows increasing use of ‘frontier AI’, both by people looking to hack IT systems and people trying to protect them. Frontier AI is not an abstract future risk, it’s the situation right now. And it’s already reshaping the economics of cyber security in ways for which most organisations are entirely unprepared.
That's why while a wider debate rages on over whether the latest generation of frontier models are inherently dangerous, the bigger issue for me is how CISOs should be addressing the very real threats that frontier AI poses.
Not only is frontier AI accelerating vulnerability discovery and compressing the time between disclosure and exploitation (to the benefit of hackers), it is also exposing exactly how vulnerable organisations are when their IT security processes are slow, periodic and reactive (to the distinct disadvantage of IT security teams).
In short, this will be a losing fight for IT teams that continue to rely on outdated practices such as static controls and fixed remediation windows. The prospects don’t look good either for those that have only fragmented visibility into anomalous behaviour on company systems or suspicious traffic flows on company networks.
On the bright side
However, there still are grounds for optimism. AI may be the biggest accelerant to cyber risk today, but it is also security's greatest potential equaliser. This is only if organisations are willing to rethink how they deploy it. Truly realising that potential requires moving beyond the idea that AI is simply a faster version of what came before.
Models at this capability level don't just raise policy questions. They raise the ceiling and expand the attack surface on what autonomous attacks can do.
The conversation has shifted. It is no longer just about AI helping analysts work more efficiently. It is about autonomous agents: AI systems that can act, adapt, and chain decisions together without human intervention at every step. Attackers are already exploiting this. Autonomous agents can probe systems continuously, adapt to defences in real time, and identify exploitable vulnerabilities at a speed and scale no human team can match. The attack surface this creates is fundamentally different from anything most organisations have planned for.
For defenders, the answer cannot be to simply run faster on the same track. Agentic AI in the hands of security teams with systems that can autonomously triage alerts, investigate anomalies, and initiate containment workflows – represents a genuine shift in defensive capability. But only if the underlying architecture supports it. Agentic workflows require continuous, high-fidelity data; they break down when built on fragmented visibility or siloed signals.
That is why the organisations best placed to respond are those that have already invested in bringing together all the logs, signals and alerts that contribute to a full, real-time picture of activity across their environment. Not because it is tidy, but because agentic AI cannot function effectively without it. An agent operating on incomplete data is not a force multiplier. It is a liability.
This also raises a point that deserves more attention from CISOs: model dependency. The events of recent weeks have demonstrated that reliance on any single frontier model creates concentration risk: operational, regulatory, and reputational. Security architectures that are designed to be model-aware, where the intelligence layer can flex without disrupting the workflows beneath it, are inherently more resilient. That is not a theoretical advantage. It is a practical one.
The sophistication of both attacks and defences will increase as AI technology advances. CISOs who treat this as a reason to wait are making a decision by default, and it is not a good one.
To summarise: if your IT security team is simply trying to patch vulnerabilities faster, it is solving the wrong problem. Vulnerability management can no longer be treated as a scheduled hygiene exercise when AI can help attackers find and act on weaknesses at machine speed. And when autonomous agents can chain those weaknesses into full attack sequences before a human analyst has even opened a ticket, the gap between unprepared and resilient organisations will only widen.
From a CISO point of view, I firmly believe that the organisations that stay resilient will not be the ones where security leaders are still debating whether frontier AI is hype or threat. They will be the ones that have already stopped asking whether this is real and started making sure their teams have what they need to respond to it.
Read more in this series
- John Bruce, Quorum Cyber: Claude Mythos forces the conversation on defensive AI.
- Martin Riley, Bridewell: Mythos is turning up the heat on risk, not rewriting the rules.
- Aditya K Sood, Aryaka: Frontier AI models could be an adversary's force multiplier.
- Ellie Hurst, Advent IM: The trust about Claude Mythos is less dramatic than it seems.
- Rik Ferguson, Forescout: What frontier AI actually means for enterprise security.
- Haris Pylarinos, Hack the Box: Why frontier AI must be stress-tested before CISOs trust it.
- Chris Atkinson, PA Consulting: AI won't break your security, but your governance might.
