The Firewall at the Edge of an AI World: Who Can Keep Up?

The Firewall at the Edge of an AI World: Who Can Keep Up?

Managing network infrastructure right now is not dull. The edge has become the most contested surface in the threat landscape, and the pace of attacks is outrunning the ability of many vendors to respond. That gap is opening up a real and growing risk, particularly for smaller organisations that cannot afford the platforms designed to close it.

The Rules Have Changed

For years, perimeter security operated on a simple premise: build a strong wall, know what is coming over it, and block what looks suspicious. Firewalls inspected traffic, enforced policy, and kept the bad out. For organisations without deep security teams, the firewall was often the only thing standing between normal operations and a serious incident.

AI has broken that model. Threat actors now use automated tools to scan internet-facing infrastructure at a scale and speed that was simply not possible a few years ago. When a new vulnerability surfaces in a firewall or VPN appliance, the window between public disclosure and active exploitation has shortened alarmingly. In documented cases from 2024 and 2025, that window was measured in hours. Not days. Hours.

The vendors paying attention have responded. Palo Alto Networks, one of the genuine leaders in this space, moved fast when CVE-2024-3400 landed: a critical zero-day in its GlobalProtect VPN feature, was actively exploited in the wild. Clear communication, an immediate workaround for customers who could not patch straight away, and a full fix within days. For teams managing these devices at scale, it was a busy period. It was also exactly the right response.

The Firewall Is Doing More Than You Think

Here is something worth understanding if you are not deep in the technical detail: an enterprise-grade firewall is not just protecting itself. It is protecting everything behind it.

Palo Alto’s Advanced Threat Prevention engine inspects every packet passing through the firewall, not by matching it against a list of known threats, but by using deep learning to analyse the intent of the code itself. That distinction matters enormously right now. A signature-based system can only catch what it has already seen. An intent-based model can identify malicious behaviour in AI-generated or polymorphic payloads that have never existed before. Even if a server sitting behind the firewall has an unpatched vulnerability, the traffic attempting to exploit it can be stopped at the perimeter before it gets anywhere near the target.

Advanced WildFire adds another layer: suspicious files are detonated in an isolated sandbox, analysed by deep learning models, and either cleared or blocked before they execute on the network. The engine is continuously trained as new threats emerge. The result is a perimeter that learns alongside the adversary, rather than perpetually chasing it.

The Scale of What the Industry Is Actually Fighting

If you want a sense of how serious this has become, look at Project Glasswing.

Anthropic launched the initiative in partnership with over 50 organisations to deploy its frontier AI model, Claude Mythos Preview, specifically for defensive security work. The coalition includes Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. Anthropic has committed up to $100 million in usage credits and $4 million in donations to open-source security organisations to back it.

The results so far: approximately 10,000 high-severity or critical vulnerabilities identified across operating systems, web browsers, and open-source projects. Anthropic deliberately chose not to release Mythos publicly precisely because of this capability. A model that can scan a codebase, identify dormant vulnerabilities, and construct working exploits is not something you put in open circulation without careful consideration.

Palo Alto has also built its own AI response from within. Its Cortex platform now includes an autonomous threat intelligence agent that continuously ingests data from CISA and NIST advisories, GitHub exploit repositories, and the broader web. An LLM reasoning engine processes that raw data, extracts the underlying attack techniques, and produces a prioritised intelligence feed for the research team. Not a list of CVEs. An analysis of exactly how each threat works and where detection coverage needs to close. When Palo Alto talks about staying a step ahead of the adversary, this is the mechanism.

One more emerging dimension: Palo Alto’s Prisma AIRS platform protects AI applications themselves. As organisations deploy internal AI tools, new attack surfaces appear: prompt injection, model poisoning, data exfiltration through AI interfaces. Prisma AIRS provides real-time enforcement across these environments. The perimeter is no longer just the network edge. It now includes the AI systems running inside the organisation.

Not Every Vendor Is Keeping Up

The vendors described above sit at the top of the market. They have the engineering depth, the threat intelligence infrastructure, and now the AI tooling to respond at the pace the current threat environment demands. Several of them are Glasswing launch partners for precisely that reason.

Below that tier, the picture looks different. There is a broad range of firewall and unified threat management products aimed at the SMB market, priced accordingly. Some come from vendors with limited security research capacity. When a vulnerability is discovered, the response timeline can be significantly longer. Weeks rather than days. Sometimes longer than that.

This is a documented problem. Both CISA and the UK’s National Cyber Security Centre publish records of known exploited vulnerabilities, and while the data consistently shows that a large proportion of successful intrusions exploit flaws for which patches have been available for months, slow vendor patching makes that problem worse. And if the most capable players in the industry are finding 10,000 critical vulnerabilities through Glasswing, the question worth asking is how many equivalent flaws exist in products from vendors running nothing comparable.

The Business That Cannot Afford the Good Stuff

This is where the risk concentrates. Enterprise-grade firewall platforms are not priced for small businesses. An appropriate deployment, with support contracts, live threat intelligence feeds, and the resource to manage it, is simply beyond the budget of most SMBs.

Those businesses often rely on lower-cost alternatives, or on consumer-grade equipment that was never designed to operate as a security boundary under active threat conditions. Some depend entirely on whatever gateway device their internet provider shipped, which may not see a security update for years.

This is not a criticism. These organisations are making rational decisions within real financial constraints, usually without dedicated security expertise. But the consequence is a large, identifiable population with perimeters that AI-assisted attacks can probe and exploit at scale, with minimal effort.

The risk does not stop with those businesses directly. Supply chain attacks use smaller organisations as a route into larger partners and customers. The weakest link does not need to be the ultimate target. It just needs to be reachable.

What Actually Helps

Managed firewall services transfer the operational burden to a provider with the scale and expertise to do it properly: configuration, patching, monitoring, and response on a timeline that reflects actual risk rather than a convenient maintenance window. For many SMBs, this is the most practical step available.

Vendor selection deserves more scrutiny than it typically gets. Patch history, response transparency, and the presence of an active security research programme are all reasonable criteria, alongside the more obvious factors of price and performance. Alignment with initiatives like Project Glasswing is a fair signal of where a vendor sits on that scale.

Cloud-based security controls, including firewall-as-a-service from providers such as Cloudflare and Zscaler, are also a credible option. They move the patching and operational responsibility to providers with significant scale, at accessible price points.

The Bottom Line

Project Glasswing is a remarkable piece of evidence. It tells us that the world’s most capable technology companies, pooling resources at a scale of $100 million, with frontier AI at the centre, are working hard to stay ahead of what AI-driven attacks can do. That is reassuring. It is also a measure of how serious the problem is.

The vendor landscape is not uniform. The best platforms are genuinely impressive. But not every vendor is at that level, and not every organisation can afford the ones that are. That gap exists, it is widening, and it will be tested.

The question for the industry is whether that gets treated as the structural risk it clearly is, before the evidence becomes too hard to ignore.

– Provided by Richard Greville – Technical Director – Opticore IT