Is cybersecurity losing the race against AI adoption?

Chief Security Officers (CSOs) are caught in a curious bind. Keen to overturn the long-standing perception of cybersecurity as a block on innovation, they increasingly strive to be strategic business enablers. Yet many are trapped by the simultaneous promise and threat posed by AI.

As businesses go all-out to drive competitive advantage through the technology, the task of charting a sensible, risk-managed course has never been more challenging. The AI attack surface is proliferating, and the old ways of doing things are no longer fit for purpose. So what happens next?

Off the starting blocks

AI use is soaring globally. According to Stanford University, adoption of generative AI (GenAI) reached 53% of the population within three years. Although this rate varies from country to country, that makes its ascendence faster than the PC or the internet.

It’s easy to see why. Whether they’re using AI to augment individual tasks, automate workflows or autonomously execute multi-stage tasks, businesses are already seeing benefits across a range of functions. Stanford cites studies reporting gains of 15% in customer support, 26% in software development and 50% in marketing output.

Yet while adoption has soared, risk management efforts haven’t kept pace. According to one study, fewer than one in ten organizations have an advanced AI security strategy.

Where risk is concentrated

They certainly need one. AI is creating new opportunities for data theft, sabotage, unauthorized actions and much more as it becomes embedded into core business workflows. GenAI chatbots pose a compliance and data leak risk so great that the Solicitors Regulation Authority (SRA) was recently forced to post a warning notice to the UK’s legal sector. A Cyberhaven study claims two-fifths of all AI interactions involve sensitive corporate data.

Then there are AI-powered web browsers, a new type of product that can be subverted by poisoned web pages or invites that abuse users’ already logged-in sessions. And agents, which Check Point Product Lead for AI Agent Security, Sam Watts, describes as a “new kind of insider threat”.

“They can go rogue on their own, as we’ve seen in recent incidents where AI systems decided to start hacking without being asked to. Or an attacker can trick or confuse an agent into acting on their behalf,” he explains. “Either way, you have a trusted insider with real access that’s been manipulated, and underneath it all sits the identities, permissions and data stores that AI can now reach.”

Two-thirds of organizations have reported at least one agent-related incident in the past year, according to the Cloud Security Alliance.

Let’s not forget either the AI stack of model servers, inference frameworks, orchestration clusters, agent control panels and other assets that sit behind these tools. In many cases it has been hastily built, leaving security holes for attackers to exploit. Of 10,000 MCP servers analyzed by Check Point AI Security Research, 40% had security weaknesses.

All of which creates a sprawling new attack surface.

“AI now lives on the device, in the cloud, across your own tech stack and inside the third-party platforms your teams rely on. External users are also using it to send content to you,” says Watts. “Previously, security teams had a fairly defined set of points to worry about. Now it spans endpoint, cloud, SaaS and partners all at once, which makes it a genuinely difficult architectural problem.”

New threats for a new attack surface

These weaknesses can be exploited by traditional methods. But the AI attack surface is also vulnerable to new techniques such as prompt injection. Here, LLMs are manipulated by malicious instructions either typed directly into the AI application or hidden in content the AI will interact with. One study scanned 1.2 billion URLs and found over 15,000 indirect-injection payloads planted in public web pages.

“We haven’t yet seen many major prompt injection attacks, because it’s something of a chicken-and-egg situation,” explains Watts. “Organisations aren’t yet comfortable deploying agents in use cases worth hacking. But as agents become a bigger part of the business, the economic incentive grows with them, and you risk a sudden shift. We already know agents are exploitable.”

Data poisoning is another AI-specific threat, this time corrupting what a model knows or retrieves rather than how it’s prompted. It can happen either at runtime, or when malicious content the model will absorb or retrieve is planted across the web. Microsoft has documented coordinated campaigns of this sort in the past.

The bad news is that shadow AI can amplify these risks. “Shadow AI is the accelerant, because it removes the one thing every control depends on: visibility,” says Chris Cochran, Field CISO and VP of AI Security at SANS Institute.

If an IT team can’t see or manage the tools being used in the organization, it increases the chances of accidental data leakage, of unpatched vulnerabilities that could be exploited, or other risks such as over-permissioned agents. One study found that nearly half (47%) of employees using GenAI in the workplace do so via personal accounts and applications.

Why legacy network security is failing

The challenge is that traditional network security is no longer effective in this new technology landscape. “It was designed around a perimeter with a small number of choke points,” says Check Point’s Watts. “AI doesn’t sit at one point. It’s on the endpoint, in the cloud, inside SaaS and third-party platforms and in agents talking directly to other systems, so there’s no single place a network control can see all of it.”

Cochran points to machine-speed AI-powered attacks as another limiting factor.

“The tempo has changed: adversaries now parse a CVE, match it to your exposed infrastructure, and generate working exploit code in hours, while your controls still run on monthly cadences,” he says. “You cannot win a machine-speed fight with a human-speed response process, and unmanaged AI usage means much of the traffic never crosses a control point at all.”

Shining a light on AI

Experts that TechFinitive spoke to agree that blocking AI isn’t the answer. CISOs should instead focus on enabling the technology safely.

“Bring AI into the light: replace bans with sanctioned tools running through a gateway you can log, and give every agent a name, an owner, and a leash,” says SANS Institute’s Cochran. “Then do the boring foundational work that becomes a superpower at machine speed: asset inventory, data classification, identity scoping, and structured logging. And push authorization itself to run at machine speed so your defenses keep pace with the threats.”

Check Point’s Watts advocates a trifecta of “discover, govern and protect”. That means understanding what’s running and then integrating “enforcement points” across the environment: pre-deployment approval processes, runtime inspection of everything going in and out of the model, and runtime protection to prevent attacks in real time.

“Underpinning all of this has to be an AI-first approach. That means up-levelling existing security tools so they work at AI speed,” he concludes. “You need AI to tackle AI. It’s only then that AI can deliver its full business value, because people can trust that using it won’t put the organization at risk.”