Last week’s decision by the US government to bar foreign nationals from accessing two of Anthropic’s most advanced AI models has reignited an important debate around who should have access to frontier AI capabilities and how they should be controlled.
The concerns are understandable. These latest-generation models represent a significant leap in capability.
But cyber security professionals have learned the same lesson repeatedly: once a capability exists, it becomes extraordinarily difficult to contain.
That does not mean governments are wrong to try. It means we should be realistic about the outcome.
Compromised access keys, increasingly sophisticated jailbreak techniques and the pace of competition across the AI industry all point in the same direction. History suggests that technological capabilities do not remain exclusive for long.
From FourNet’s perspective, this is not an abstract debate. We help organisations modernise and protect the services people depend on, bringing cyber, network and AI together under one accountable partner. Our Security Operations Centre teams see first-hand how quickly threats evolve and how rapidly new technologies are adopted.
Organisations should therefore assume that the most advanced AI capabilities available to defenders today will, sooner or later, become available to attackers as well. That should concern us.
But perhaps not for the reasons many people assume.
The biggest mistake would be to conclude that the answer is to restrict the use of AI by security teams. In fact, the opposite is true.
Defenders need these technologies just as much as attackers do.
The volume and speed of modern cyber threats have reached a point where human analysts alone cannot keep pace. AI gives defenders something they desperately need: scale.
It can help analysts investigate incidents faster, automate repetitive tasks and identify patterns that would otherwise take much longer to detect. Used properly, AI enhances human judgement rather than replacing it.
At the same time, cyber criminals understand the benefits too.
There are now dozens of highly capable models available globally. Some are commercial, some are open source, and more appear every month. Offensive uses of AI are inevitable.
We should assume that attacks powered by these technologies are coming.
Phishing emails will become more convincing. Social engineering attacks will become more personalised. Adversaries will automate reconnaissance and refine attacks at a pace that previously required large teams and significant resources.
None of this means organisations should panic.
Cyber security has always been an arms race. Attackers adopted automation. Defenders responded. Attackers embraced cloud infrastructure. Defenders adapted. Attackers are adopting AI. Defenders must do the same.
In reality, that process is already under way.
AI is not only being used by security teams. Vendors are increasingly using it to build resilience directly into networks, firewalls and cloud environments. Detection and prevention capabilities are becoming faster and more adaptive.
That matters because the pace of vulnerability discovery is accelerating. The idea that organisations can patch once a month and assume they are protected belongs to a different era.
As vulnerabilities emerge faster, organisations will need to update and respond more frequently. Vendors are already building increasingly intelligent capabilities designed to prevent attackers using networks as pivot points to compromise wider environments.
In other words, the AI race is not just happening in the hands of attackers and defenders. It is increasingly being built into the infrastructure itself.
Which brings us back to the debate around frontier AI.
The question is no longer whether increasingly capable AI models should exist. That question has already been answered.
Nor is it realistic to assume that these capabilities will remain exclusive indefinitely.
The real question is whether defenders will be equally equipped.
Holding back the good guys in the hope that the bad guys will somehow be denied access is unlikely to succeed.
Instead, organisations should focus on securing their journey to AI, building resilience and ensuring defenders have access to the same capabilities as those trying to attack them.
At FourNet, we see organisations coming through the next phase strongest when they stop treating cyber, network and AI as separate problems and start running them as one.
Good security has never been about technology alone. It has always been about people, processes and technology working in concert — and in an AI era, that principle matters more, not less.
The AI security race has already started. The question is no longer whether increasingly capable models will proliferate. History suggests they will.
The real question is whether defenders will be equally equipped.
Organisations that treat AI, cyber and network resilience as one challenge rather than three separate ones will be better placed to secure the services people depend on.
Because in cyber security, resilience beats wishful thinking every time.
