AI Just Killed Security Through Obscurity. It Was Never That Healthy to Begin With.
Security through obscurity was always a bad idea. The notion that systems stay safe because attackers simply don't know enough about them has never been a sound foundation for anything. It was the security equivalent of hoping nobody tries the door. And yet, plenty of organisations leaned on it anyway, through a combination of limited resources and wishful thinking.
That door is now wide open.
AI agents are being used by researchers, vendors, and frankly criminals, to tear through codebases looking for vulnerabilities. Old code. Obscure code. Libraries that have been quietly running inside 80 percent of web servers for a decade while the community collectively assumed someone else had already checked them. Brett Leatherman, assistant director of the FBI's Cyber Division, confirmed what many privately suspected: the latest models are finding significant vulnerabilities in software that the community believed had been thoroughly stress-tested. Believed being the operative word.
The scale of this is visible in the patch numbers. Microsoft's most recent Patch Tuesday addressed 974 CVEs in a single drop, which is not a typo. Dustin Childs, chief bug hunter at Trend Micro's Zero Day Initiative, noted that the patched components included things like the Telnet client, Windows RNDIS, NFS Portmapper, and the Vista-era Link Layer Topology Discovery protocol. Components nobody has seriously thought about in years, quietly sitting there, full of problems.
The attacker side of this equation is equally grim. AI is being used to reverse-engineer patches and develop working exploits within hours of a fix being published. In one recent incident, at least four espionage groups with suspected Chinese links exploited an open-source Chromium vulnerability using a kit developed in the window between an upstream patch being released and the stable downstream version reaching users. That window used to be a relatively safe buffer. It is not anymore.
OT Security: The Part That Should Actually Worry You
Operational technology is where this gets serious. Industrial control systems, water treatment, energy infrastructure, manufacturing, the stuff that keeps the physical world running, have historically been somewhat insulated by the sheer obscurity of their protocols and proprietary hardware. Attackers generally needed specialist knowledge to cause damage. That acted as an informal barrier to entry.
AI removes that barrier entirely.
John Hultquist, chief analyst at Google Threat Intelligence Group, described the problem plainly: AI is excellent at navigating obscure systems and providing technical guidance, which makes it a powerful tool for anyone targeting infrastructure they would otherwise know nothing about. 'They've been largely secured because the expertise was in a handful of people's heads,' he said, 'and that's not going to last forever.'
This is not hypothetical hand-wringing. Five US agencies recently confirmed that attackers used AI-generated exploitation scripts to compromise internet-exposed Siemens S7 Series programmable logic controllers at water, energy, and manufacturing facilities. Real attacks. Real infrastructure. The feds described it as 'an active threat,' which in government language is about as blunt as it gets.
AI can assist attackers at every layer, from the application down through the operating system and into firmware. For legacy industrial systems, many of which were never designed with internet exposure in mind, this is a genuinely serious development.
Finding Bugs Was Never the Hard Part
Katie Moussouris, founder of Luta Security and someone who has spent more time thinking about vulnerability disclosure than most, made a point worth sitting with: finding bugs has never actually been the hard part. Triaging them, prioritising them, and getting them fixed, that is where organisations consistently fail.
AI is currently very good at accelerating the discovery side. The defensive side is another matter.
Research from 1Password tested two frontier models across six recent CVEs, generating over 6,000 patch attempts. The average success rate for a patch that actually fixed the vulnerability without breaking anything else: 26 percent. More than half the time, the AI-generated patch either failed to fix the problem, introduced a new one, or managed both simultaneously.
Veracode ran a broader study across more than 100 models and 80 coding tasks, finding that AI-generated code passed security checks just 56 percent of the time. That is not a pass rate anyone should be comfortable with near production systems.
So the pipeline looks something like this: AI finds vulnerabilities faster than ever, attackers exploit them faster than ever, and the tools meant to fix them are wrong more often than not. The treadmill is speeding up while the trainers are falling apart.
Moussouris is particularly critical of organisations that respond to this situation by simply throwing more resources at finding and patching bugs without examining why they keep producing so many in the first place. Pattern recognition matters here. If your codebase keeps generating injection flaws, the answer is not to patch them faster, it is to fix the part of your development process that keeps producing them. Better templates earlier in the CI/CD pipeline prevent vulnerabilities at scale rather than playing whack-a-mole indefinitely.
Counting bugs fixed and measuring patch speed are useful metrics, but they tell an incomplete story. The organisations that will survive this period are the ones that understand what their vulnerability patterns say about their underlying processes, and actually change those processes. The ones that just run faster on the same broken treadmill are going to exhaust themselves.