C-urity had a black great October: AI exposes flaws in traditional cybersecurity models

October witnessed a surge in both cybersecurity threats and the stock prices of cybersecurity firms, as companies grappled with the limitations of traditional security strategies. The rise of AI is prompting a reevaluation of data protection appro...

ET Bureau

AI-powered cyberattacks expose vulnerabilities in enterprise software, highlighting the need to minimise sensitive data exposure rather than relying solely on threat detection and response.

October was a terrible month for cybersecurity. AI-powered attacks exposed vulnerabilities across widely used enterprise software, attackers compromised systems at unprecedented speeds, and security researchers struggled to keep up. Yet, stocks of cybersecurity companies rose, enterprises continued to increase security budgets, and the industry found itself with more demand than ever.

The reason for this growth lies in a fundamental assumption that has shaped cybersecurity for two decades - one that AI may force us to reconsider. Large companies spend heavily on cybersecurity products to detect threats, monitor systems, identify vulnerabilities and respond to incidents. When a breach occurs, the natural response is to spend more on better detection, monitoring and incident response.

This isn't irrational. These products serve a purpose. But the cybersecurity industry operates on the assumption that sensitive data is spread across thousands of applications, databases, cloud services and, now, AI systems. Security, therefore, is about protecting these systems and detecting attackers.


pic1
<p> Trapped by complexity<br></p>
Consider a bank with customer data in its core banking system. The same data gets copied into its CRM, analytics platforms, marketing systems, fraud detection applications, data warehouses and dozens of SaaS products. Each decision to copy the data makes perfect sense in isolation. But taken together, these decisions create a rather unreasonable security architecture.

Every copy becomes another place where sensitive data can be stolen, another system that needs access controls, and another potential breach. Yet, we've built an enormous cybersecurity industry around protecting this architecture rather than questioning it.

AI is now exposing limitations of this approach. In October, Google's threat intelligence team described an attacker going from compromising a cloud resource to running an AI-generated credential-harvesting campaign in under 6 hrs.
ADVERTISEMENT

The underlying weaknesses weren't particularly novel. Exposed credentials, excessive permissions and vulnerable software have been with us for decades. What AI changes is the speed and cost of discovering and exploiting them.

pic2
<p>Systems within systems<br></p>
The industry's response is predictable: use AI to improve detection and response. That's reasonable, but there's a problem. If attackers and defenders both get access to increasingly capable AI models, why should we assume defenders will stay ahead? More importantly, why is our entire security strategy so dependent on winning that race? What if, instead of trying to protect thousands of copies of sensitive data, we reduced the number of copies in the first place?

A payment application might need to process a credit card, but that doesn't mean every system involved needs to store the card number. Similarly, an AI agent may need to perform a task involving sensitive data without requiring unrestricted access to the underlying information.

Technologies like tokenisation, encryption and data isolation make this possible. Rather than distributing sensitive data everywhere, applications can operate on safe representations, with access to original data governed by explicit policies. If an attacker compromises an application, they don't automatically get access to the sensitive customer data that would traditionally have been stored there.
ADVERTISEMENT

This becomes particularly important with AI agents. Unlike traditional applications, agents can dynamically decide which tools to use, which systems to call and what actions to take.

Their usefulness comes from operating across multiple systems, but that also makes their access to sensitive data increasingly difficult to control. We cannot rely exclusively on monitoring unpredictable AI systems. We need deterministic controls over what data they can access, regardless of what they decide to do.
ADVERTISEMENT

In some ways, AI agents are exposing a problem that has existed in enterprise software for decades. We've confused an application's need to use data with its need to possess that data. Those are not the same thing.

None of this means firewalls, endpoint security, threat detection or incident response are going away. We will always need them. But perhaps we should spend as much energy reducing the amount of sensitive data exposed across the enterprise as we do monitoring all places where it exists.

The irony of Black October is that the cybersecurity industry may be celebrating another period of strong growth while its customers face an increasingly difficult security problem.

It's time we measured the success of cybersecurity not just by how quickly we detect and respond to breaches, but by how little an attacker can steal when one happens.

The writer is co-founder-CEO, Skyflow.
(Disclaimer: The opinions expressed in this column are that of the writer. The facts and opinions expressed here do not reflect the views of www.economictimes.com.)
Download
The Economic Times Business News App
for the Latest News in Business, Sensex, Stock Market Updates & More.
READ MORE
ADVERTISEMENT

READ MORE:

LOGIN & CLAIM

50 TIMESPOINTS

More from our Partners

Loading next story
Business News › Opinion › ET Commentary › C-urity had a black great October: AI exposes flaws in traditional cybersecurity models
Text Size:AAA
Success
This article has been saved

*

+