Twenty-seven years. That’s how long a critical security vulnerability sat undetected inside OpenBSD — one of the most security-hardened operating systems ever built. An AI found it. And that bug was just one of thousands.
Anthropic’s most capable model, Claude Mythos, was recently aimed at the operating systems and browsers the world runs on. It found thousands of previously unknown zero-day vulnerabilities — flaws no human had documented, never been exploited, never been patched. The model is so capable — and so dual-use — that Anthropic won’t release it publicly. Instead, they launched Project Glasswing: a controlled program giving Microsoft, Google, Apple, Amazon, and the Linux Foundation access to use it for defensive security work.
If you’re an accountant or finance professional, you might be thinking: that’s a tech story, not my story. I’d push back on that. What just changed is the speed and scale of software risk — and software risk, in 2025, is financial risk, operational risk, and reputational risk for nearly every client and organization you work with.
How Vulnerability Management Has Always Worked
Think about it this way: we’ve been building the digital infrastructure of the global economy out of sticks. Handcrafted, not load-tested, shaped by whoever was working that day. AI is the first industrial-scale lumber mill.
Historically, vulnerability management ran on a slow, human-paced cycle. A researcher would find a bug, notify the vendor (usually giving them around 90 days), the vendor would patch it — which could take weeks, months, or years — and then push the update to users who may or may not install it. The main bottleneck was always the human: one person, choosing what to look at, spending time finding the flaw.
That bottleneck is gone.
The Kaseya attack on July 3rd, 2021 is a useful case study. Researchers had found a vulnerability in remote management software months earlier and notified the vendor. The vendor patched two of three related issues — but the third was still open when a ransomware gang found the same bug over the July 4th weekend. Thousands of organizations compromised. Grocery stores shut down. Credit unions offline. All because of a patch that was in progress but not finished.
There’s also a quieter version of this problem that’s been building for years. In 2013, Microsoft’s internal bug-tracking database — their master list of every known unpatched vulnerability — was hacked. The public didn’t find out until 2017. Tech companies had no disclosure obligation. These hidden exposures have been accumulating in the software stack the global economy depends on, largely invisible to the people who rely on it.
What AI Changed
AI can now scan millions of lines of code across entire operating systems autonomously. It can recognize how separate flaws interact and chain them together into larger exploit paths. It can construct working attack code — not just identify the flaw. And it can do all of this on both sides: defending and attacking.
The race isn’t AI versus no AI. It’s defensive AI versus offensive AI. On the defensive side, Project Glasswing gave major tech companies controlled access and $100 million in credits to open source security organizations. On the offensive side, hacker-specific AI tools are available on the dark web right now for as little as $50 on a lifetime subscription.
The defensive side currently has the better tool. That lead is not permanent.
What This Means for Accountants and Finance Professionals
Third-party and vendor risk assessments just got harder. The standard question — “are you current on your patches?” — is no longer a sufficient proxy for security posture. The volume of newly discovered vulnerabilities is overwhelming patch pipelines at even the largest vendors. The right question now is: what’s your patch velocity, and how do you prioritize critical zero-days as they emerge? That question will tell you a lot about how seriously a vendor treats operational risk.
Internal controls conversations need to include software update posture. The gap between when a patch becomes available and when active exploitation begins is compressing fast. In the Kaseya case it was months. In a more recent Exchange vulnerability case, one organization waited six hours to test the patch — and was already compromised when they applied it. If your firm or your clients are running critical systems on quarterly patch cycles, that posture may no longer be defensible.
There’s also a systemic risk angle worth naming. When hundreds of millions of people run the same operating system or enterprise platform, a single unpatched vulnerability becomes global exposure. This is the software monoculture problem — and it has a direct parallel to counterparty concentration risk. If your clients or your firm are heavily dependent on the same software stack as millions of others, a single flaw is a shared exposure.
Key Takeaways
- The patch gap is growing. AI discovers vulnerabilities faster than humans can fix them — and that asymmetry is a risk management problem, not just a tech problem.
- The race is defensive AI versus offensive AI. The defensive side has the better tool right now. That won’t last forever.
- “Are you patched?” is no longer sufficient in vendor due diligence. Ask about patch velocity and zero-day response protocols.
- Software monoculture is the cyber equivalent of counterparty concentration risk. Shared software dependency = shared exposure.
- Update your software. When the prompt appears — not this weekend.
Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: AI Found Thousands of Zero-Days
Leave a Reply