AI in cybersecurity has crossed a line most security teams didn’t expect this soon. Machines are now finding bugs that skilled humans missed for more than a decade. Yet the same AI that uncovers a hidden flaw can write a patch that quietly opens a new one.
That tension sits at the centre of a recent BusinessLine report. Among the experts it consulted was Anuj Magazine, Co-Founder and CTO of AI&Beyond, the AI literacy and adoption firm he leads with co-founder Jaspreet Bindra. His view is clear: AI is a powerful new lens for security, but it is not a replacement for human judgment.
Here is what that means for business and technology leaders.
The wake-up call: a decade-old flaw in PHP
On October 21, 2026, PHP shipped fixes for two critical vulnerabilities. Both could allow unauthenticated remote code execution on web-facing installations. Users were told to move to the latest supported PHP 8 versions.
The striking detail is how long the bugs had survived. They sat in PHP’s codebase for more than ten years. Cloud security firm Wiz found one of them using its autonomous research agent, Wiz Atlas, working alongside Google’s Gemini 4 Argon and OpenAI’s GPT-5.6 Sol.
This is AI in cybersecurity at its most impressive. It is also a preview of the risk.
How AI sees what traditional tools miss
Anuj Magazine explains the shift in simple terms. Conventional security tools mostly hunt for known patterns of mistakes. Others use fuzzing, throwing huge volumes of random input at a program to see what breaks.
AI can go a step further, he says. It can study how different parts of a system connect and spot unusual combinations of conditions that create risk. Those are exactly the flaws that rule-based scanners overlook, because no single line of code looks wrong on its own.
The evidence is growing. In a 2025 DARPA contest, AI systems found 18 previously unknown vulnerabilities in real open-source software and submitted 11 patches. Zero-day vulnerabilities, the weaknesses nobody has fixed yet, are where AI vulnerability detection shows the most promise.
Anuj Magazine’s warning: one fix can open another gap
This is where AI&Beyond’s perspective becomes most useful. Finding a flaw is only half the job, and Magazine is cautious about the other half.
First, prediction is hard. AI can help rank which known weaknesses attackers are likely to target. It cannot reliably foresee every new attack technique.
Second, and more important, he notes that AI “can also close one security gap and unintentionally open another.”
The data backs him up. A 2026 internal study by 1Password reviewed more than 6,000 AI-generated security fixes across six complex vulnerabilities. Only 26% fully fixed the problem without changing how the software behaved. In roughly 54% of cases, the fix failed, introduced a new weakness, or did both.
In other words, AI-generated security patches fail more often than they fully succeed. Teams that ship them unchecked are trading a known risk for an unknown one.
The attacker’s advantage is shrinking the patch window
There is a second problem. AI in cybersecurity is not only a defender’s tool.
Amazon detected an attempt to exploit the PHP flaws on October 17, days before the public fix. The company warned that AI-assisted analysis is cutting the cost of turning vulnerability details into working exploits. More attackers can now build one within hours, sometimes minutes, of disclosure.
That changes the maths for every organisation. The comfortable gap between disclosure and attack is closing. Quarterly patch cycles no longer match the speed of the threat.
What the experts agree on: AI plus human oversight
Sachhin Gajjaer, Founder and CEO of Sattrix Group, adds an operational view. AI agents can scan code, suggest fixes and run automated tests. But a wrong patch can create fresh vulnerabilities or break the application, so people must still review and validate each change.
The experts in the report reach the same conclusion. AI should complement conventional security testing, not replace it. Human oversight in AI security is what filters false positives and stops a fix from becoming the next breach.
Why this is really an AI literacy problem
This is the deeper lesson behind AI&Beyond’s stance. The tools are already powerful. The gap is in how people use them.
A developer who trusts an AI patch blindly is not facing a technology failure. It is a literacy and adoption failure. Teams need to know what AI is good at, where it breaks, and when to step in. That is the gap Anuj Magazine and Jaspreet Bindra built AI&Beyond to close.
For leaders, a practical playbook looks like this:
- Use AI to widen discovery. Point AI tools at legacy and long-lived code, where hidden flaws are most likely.
- Treat every AI patch as a draft. Require human review and regression testing before deployment.
- Shorten patch cycles. Assume exploits can follow disclosure within hours.
- Measure outcomes, not output. Track how many AI fixes fully resolve issues without side effects.
- Train people, not just tools. Build AI literacy so engineers know when to trust and when to question.
The bottom line
AI in cybersecurity is no longer experimental. It is finding real, critical flaws faster than humans can. But speed without judgment is its own vulnerability.
As Anuj Magazine’s view makes clear, the winners will not be the firms with the most AI. They will be the firms whose people know exactly how far to trust it.