Artificial Intelligence & Machine Learning
,
Governance & Risk Management
,
Legacy Infrastructure Security
Advanced AI Models Find More Holes Than Enterprise Security Teams Can Plug
When Project Glasswing was announced last month, Anthropic disclosed that its most advanced frontier artificial intelligence model had found vulnerabilities in open-source software that were 16 and 27 years old, and Claude Mythos is continuing to up the stakes for software developers.
See Also: Know Thy Enemy: Threats to Cyber Resilience
Recently Palo Alto Networks, a member of the Glasswing coalition, ran Mythos against its code base, and its findings revealed that “these models are likely even better at finding vulnerabilities than we initially realized,” wrote Lee Klarich, CPTO, Palo Alto Networks, in a blog post. After scanning 130 products, the company’s May “Patch Wednesday” security advisory disclosed 26 CVEs covering 75 vulnerabilities. In a typical month prior to using Mythos, they released five.
It was the first time the majority of the company’s findings were the result of a frontier model code scanner, and it won’t be the last. “It’s important to understand this isn’t a one-and-done situation,” Klarich wrote. “We’re now rescanning, applying all our learnings about how to provide the right context and threat intelligence to the models.”
Security vulnerabilities have long been contributors to legacy systems technical debt. How deep will the tech debt hole go now that powerful AI models are doing the searching?
Tech Debt All the Way Down
Tech debt has existed for as long as people have been writing code, but the speed and precision with which AI is finding vulnerabilities are rapidly transforming the landscape, turning a new spotlight onto an old problem that has lurked in the IT shadows for decades.
Christopher Frenz, CISO at Rectangle Health, said the problem is the result of years of misguided strategy.
“The approach security has traditionally taken is the wrong approach,” Frenz said. “It’s always been focused heavily on reaction rather than being proactive. Security teams leave failure modes in place. They’re scared to disable legacy protocols, because what if somebody needs it in the business? There is security debt present in the organization because teams are afraid to ever take things away.”
Frenz said this reluctance keeps security teams chasing the wrong problems, but that AI is forcing them to reevaluate their approaches.
“A lot of organizations have a bunch of leaky pipes, and rather than ever fixing the pipe – which would be taking away the attack path – what they do is they keep buying more leak detectors and bigger buckets. They never really solve the problem. They just get better and better ways to identify the fact that problems exist.”
Erik Nost, principal analyst at Forrester, said that Mythos has enterprise technology leaders realizing that they could have a bigger tech debt problem than they’re ready to manage. Clients have been asking him if Mythos is real or if the fervor around it is all marketing hype, and his answer is that it’s a little bit of both. “But that’s not to discount the validity of the advancements that Mythos has had,” he said.
Large companies with volumes of software are taking a multi-model approach, whether that’s using Mythos, Opus 4.7, GPT 5.5 or other advanced models, depending on their access and vendor ecosystem, he said.
But the bigger problem, Nost said, is a general lack of organizational readiness to manage the pace. While the fundamentals of vulnerability management – visibility, prioritization and remediation – remain the same, AI is turning up the dial.
“With AI models leading vulnerability discovery and exploit chaining, it gets faster,” he said. “The reality is organizations just aren’t ready for that. We’re still working in legacy processes, we still have legacy code. This is the shift that teams need to prepare for.”
In late 2024, a Forrester survey found that 75% of technology decision makers expected tech debt to reach “moderate or high” severity levels in 2026, driven by the rapid expansion of AI.
More recent data shows that the problem is even worse. A Veracode study released in February 2026 found that security debt now affects 82% of organizations, up from 74% a year ago and critical security debt affects 60% of organizations, a 20% increase over the same period.
When AI discovery tools can potentially surface many vulnerabilities from a single code base scan, organizations can easily be overwhelmed.
“One organization I talked to that has been running these models recognizes that they do not have the capacity from their dev team to fix all the vulnerabilities that are being found,” Nost said. “They’re looking at preventative controls – like virtual patching – to decrease the likelihood that those vulnerabilities might be exploited.”
But that approach has limits, he said, and even web application firewalls can come with new security flaws. “Sometimes introducing these controls can add not just availability risks, but potentially security risk as well,” Nost said.
Some tech leaders are even considering the previously unthinkable and prioritizing security over availability, he said. “Now there’s discourse at some of these companies with leaders around: When do we accept the risk of downtime over the risk of a potential breach due to a vulnerability discovered with a model? These conversations are starting to happen at a high level.”
The tech debt problem is compounded by the fact that AI is both finding and creating tech debt as more development teams use AI to write new code at a pace that is overwhelming for security teams. “I think for the short term, AI is going to compound the tech debt problem, because it’s going to make things vulnerable faster than teams can eliminate it,” Frenz said.
What CIOs and CISOs Can Actually Do
For enterprises looking to mitigate the tech debt onslaught, Frenz said he recommends an architecture-first approach: figure out what you need to do business and remove everything else, taking away potential paths for attackers to exploit.
“If you analyze most organizations, you’ll find that a lot of the paths that exist through the organization are actually never used. You can actually remove the functionality and it’s not going to impact the business in any way, because no one’s actually using it,” he said.
Frenz implemented zero trust architecture at an underfunded safety-net hospital in 2015 – proof, he said, that the approach is achievable under real-world constraints, not just in greenfield environments with unlimited budgets.
He highlighted the Log4J vulnerability as a useful example.
“Log4J was a very famous vulnerability, very widespread, impacted a lot of organizations,” Frenz said. “But if you were able to break the attack path by having a control like egress filtering in place, you didn’t have the outbound communication necessary for the exploit to actually work in an unpacked state. You could keep yourself protected because the attack path is broken at a different point. Even if the vulnerability exists, it becomes essentially non-exploitable, because the path to take advantage of it isn’t there anymore.”
Frenz uses a metric he calls the Path Erasure Rate, which measures the total number of attack paths in an environment against the number that a given action would eliminate, to prioritize action. “Whatever changes are going to give you the biggest reduction in attack paths – that’s the way to go. The biggest delta is what I’m looking for,” he said.
Ultimately, the same technology surfacing decades of hidden debt may be what pays it down.
When Mozilla ran Mythos, it found hundreds of bugs that it said couldn’t have been remediated without AI. “There’s this other side where AI is accelerating and helping security teams group these things together and come up with better remediation plans,” Nost said. “We run these models, we find the vulnerabilities before we commit and we use AI in our dev tools to fix them.”