Opinion: The Patch Tsunami, Six Months In
By Jason Garbis, Numberline Security Founder and CEO
Assisted, as always, by Scully the Terrier, who has strong opinions about treats, and none whatsoever about coordinated vulnerability disclosure.
Back in April, I wrote about the coming patch tsunami. This was the idea that AI-assisted vulnerability discovery, epitomized by Anthropic’s Claude Mythos and similar models, would find security flaws faster than the industry could realistically fix them. I said the growing discovery-to-patch gap was the real story, not the discovery itself.
Six months in, I wish I could tell you I was wrong.
I’m not. The numbers, where they exist, are stark: across 281 open-source projects, Anthropic has disclosed 1,596 vulnerabilities, of which 75 have been patched with public advisories. Do that math with me: that’s under 5%. And that’s just the open-source slice. The partner-reported total across the full Glasswing coalition exceeds 10,000 high- and critical-severity findings, the overwhelming majority of which remain under embargo, unresolved, or simply waiting in a queue that was never built for this volume. One more data point: Microsoft’s September Patch Tuesday included 974 CVEs.
This isn’t a failure of the maintainers, or of Anthropic, or of anyone in particular. Anthropic’s own May update put it plainly: progress on software security used to be limited by how quickly new vulnerabilities could be found, and now it’s limited by how quickly they can be verified, disclosed, and patched. That’s exactly the bottleneck I described in April, now measured instead of predicted.
A few developments since April are worth your attention directly:
19 Days Without Access
The coalition tripled, then the models briefly vanished. Project Glasswing expanded from roughly 50 founding partners to about 150 additional organizations across more than 15 countries by early June, bringing the coalition to around 200 total. Then, on June 12, the U.S. government issued an export control directive ordering Anthropic to suspend all access to Mythos 5 and its public sibling Fable 5 for any foreign national, worldwide. Because Anthropic had no reliable way to verify user nationality in real time, both models were disabled for every customer, everywhere.
The suspension lasted 19 days. Export controls were lifted June 30, and Anthropic began restoring global access to Fable 5 on July 1. Mythos 5 itself, notably, remains restricted to a set of vetted U.S. organizations even now.
I want to sit with that for a second, because I think it’s the most important governance signal to come out of this whole saga. We built a tool intended to give the good guys a head start. That tool got tangled in export control policy so fast it disappeared for nearly three weeks, with no advance warning to the enterprises depending on it. If your security strategy assumed continuous access to any single AI vendor’s capability, June was a preview of what happens when that assumption breaks. Plan accordingly, and strongly consider locally-deployed open weight models when possible.
The Missing 90 Day Report
A 90-day report was expected in early August. As of this writing, we haven’t seen it land. Anthropic committed to a formal public summary covering fixed vulnerabilities and operational learnings roughly 90 days after the June expansion. We’ll be watching for it, and we’ll cover it here the moment it’s out. But its absence so far is its own small data point: even the report about the patching bottleneck is running behind schedule.
Why the Vulnerability Patch Gap is Measured in Months
None of this changes my prescription from April. If anything, it sharpens it. Patching remains necessary, but it cannot be your primary defense when the discovery-to-patch ratio looks like this. Default-deny, Zero Trust-based access control is what buys you safety in the gap between “vulnerability exists” and “vulnerability is fixed.” A gap that the data now shows is measured in months, not days, for the overwhelming majority of findings.
The tsunami didn’t arrive as a single wave. It’s arriving as a slow, grinding tide, and six months in, the tide is still coming in faster than the seawalls are going up. Even the AI firms’ leaders are reassessing this pace, and asking to slow things down.
The patch queue isn’t getting shorter. Your exposure can.
Book a free 30-minute Zero Trust briefing. We’ll look at where your unpatched systems are most exposed and how the Zero Trust Blueprint closes that gap.