AI Is Exposing Cybersecurity's Biggest Assumptions

Artificial intelligence is rapidly becoming the center of the cybersecurity conversation.

Every week, there seems to be a new AI-powered analyst, AI SOC, AI agent, or AI assistant promising to transform cybersecurity.

The assumption is that intelligence is the missing ingredient. That if we add enough AI, security teams will finally be able to keep up with the scale and complexity of modern attacks.

I think the industry is solving the wrong problem. René Bonvanie and I discussed this in the latest episode of Inside Cybersecurity, where we explored why AI is exposing limitations in the architectures cybersecurity has relied on for years.

An AI agent cannot defend what it cannot see. And most cybersecurity systems were never designed to provide the complete visibility that modern AI requires. Before we talk about smarter models, we should talk about the architecture and data foundations those models depend on.

AI Changes the Requirements

For the last two decades, cybersecurity has been shaped by a series of reasonable decisions. Storage was expensive, compute was expensive, and moving large volumes of data was difficult. Security products were designed to collect the information they needed, analyze it, and move on.

Those decisions made sense, for that era. However, artificial intelligence changes the equation.

Unlike traditional security tools, AI improves with more data, more context, and more history. The less information an AI system has access to, the less effective it becomes. AI does not reduce the need for data. It dramatically increases it.

That creates a challenge because most cybersecurity architectures were not designed with those requirements in mind.

The Visibility Problem

Over time, cybersecurity products expanded the amount of information they collect and analyze. Logs and events were joined by telemetry from endpoints, networks, cloud environments, SaaS applications, identity systems, and dozens of other sources.

The problem is that the data remained fragmented.

Every product collects its own telemetry. Every platform stores its own data. Every vendor claims to have a data lake. In reality, most organizations operate dozens of separate data stores (I call them “puddles”) that rarely communicate with one another.

I have written previously about both data and sovereignty. These concepts become significantly more important in the AI era because AI depends on complete visibility into an environment. Without that visibility, even the most advanced models operate with blind spots.

Financial institutions, governments, critical infrastructure providers, healthcare organizations, and defense agencies often operate under strict regulatory, privacy, and security requirements. Many cannot simply move their most sensitive information into a public cloud environment and allow third-party AI systems to operate on it. These organizations have less flexibility than ever before in how that data can be used.

The Economics of AI Detection

LLMs are impressive technologies. They can classify content, summarize incidents, explain findings, generate reports, and assist security analysts. They will absolutely have a role to play in cybersecurity.

Cybersecurity operates at a scale that most discussions about AI tend to ignore.

Let’s assume, highly optimistically, that an organization can somehow identify the correct 1,000 security events, every second, that should be analyzed by a large language model. That assumption alone is problematic because identifying the important events is often the detection challenge itself.

Now let’s assume those events are only 100 bytes each. At roughly four bytes per token, that produces approximately 25,000 tokens every second.

Here we run into the need for history when detecting attacks, since LLMs are not very good at remembering. If we use a one-million-token context window, we can analyze roughly 40 seconds of activity. If we assume an aggressive 5:1 compression ratio, which introduces its own risks in a cybersecurity environment, we can extend that to approximately 200 seconds. For this example, we assume a single one-million-token inference is performed every second using a rolling 200-second lookback window. At that point, the economics become straightforward. The annual cost of operating an LLM detection system is driven by four variables:

Annual Cost = Tokens in the Context Window × Queries per Second × Seconds per Year × Cost per Million Tokens

For simplicity, we will ignore the cost of the internal and output tokens. Using current Claude Opus pricing, continuously operating at that scale would cost approximately $158 million per year.

Cost is only part of the challenge. Large context windows also introduce significant prefill latency. Depending on model implementation, cache behavior, and service load, processing a one-million-token context can take tens of seconds, or even several minutes, before a response begins. For many real-time detection workflows, that latency is difficult to accommodate.

And this is the optimistic scenario.

It assumes we somehow selected the right events. It assumes the events are relatively small. It assumes compression does not remove important security contexts. And it assumes the attack can be detected within a very limited observation window.

Can the cost be reduced? Of course. We can analyze less data. We can shorten the context window. We can query less frequently. But each optimization introduces another compromise.

Detection and Explanation Are Different Problems

This is where the discussion around AI often becomes muddled. Detection and explanation are very different problems.

Machine learning has been helping security teams identify anomalies, correlate events, and recognize attack patterns for more than a decade. These systems are designed to operate on large volumes of data and identify signals that would be impossible for a human analyst to find manually.

Large language models solve a different problem. Once suspicious activity has been identified, an LLM can help explain what happened. It can summarize findings, generate reports, recommend actions, and accelerate investigations. These are valuable capabilities.

But they should not be confused with the detection layer itself.

• Adding an AI assistant to a security platform does not solve a visibility problem.

• Adding an agent does not solve a data problem.

• Adding a chatbot does not solve an architecture problem.

If the underlying system operates on fragmented data, incomplete context, or constrained visibility, AI can only work with what it can see. The quality of the outcome will always be limited by the quality and completeness of the underlying data.

What Comes Next

This does not mean the cloud is broken. It does not mean AI is overhyped. And it does not mean organizations should abandon the architectures that brought us this far.

It does mean that AI is exposing assumptions on which many cybersecurity architectures were built. The systems we built over the last twenty years were designed for a different set of constraints. The AI era introduces new ones.

Organizations now need more data, more context, and more visibility than ever before. At the same time, many face increasing regulatory, privacy, operational, and sovereignty requirements that limit how that data can be collected, processed, and shared.

Those competing forces are creating new challenges for cybersecurity teams.

Over the coming months, we’ll explore those challenges in greater detail. We’ll examine the compromises organizations make to manage security data, the economic realities that shape cybersecurity architectures, why building these systems internally is harder than it sounds, and what a different approach could look like.

The future of cybersecurity will be determined by whether the systems beneath that AI are capable of supporting it.

— Nir