AI is finding bugs faster than most businesses can patch them
Between April and July 2026, Anthropic’s Project Glasswing initiative identified at least 129,000 verified software vulnerabilities. More than 33,000 of those were rated critical or high severity. And Anthropic itself says that number is probably an undercount, estimating the true impact could be at least five times higher. If you run a small or mid-size business in New Jersey, that’s not an abstract research statistic. That’s your software stack.
The practical consequence is straightforward: AI is now finding vulnerabilities at a pace that manual security research never could. That’s good for defenders in theory, but it also means the window between “flaw discovered” and “flaw exploited” is getting shorter. Your patch process needs to keep up, or you’re exposed longer than you used to be.
What Anthropic actually built, and why it matters
Anthropic announced on October 7, 2026 that it’s expanding its Cyber Verification Program, which gives vetted security teams access to its Claude AI models with reduced safeguards for legitimate defensive and offensive security work. There are three tiers: Defense Access for incident response and vulnerability analysis, Red Team Access for authorized penetration testing, and Specialized Access for a small group of organizations testing safety systems themselves.
The models available include Claude Opus 5.5 and Claude Sonnet 5.5, among others. In benchmarks, the Red Team Access tier on Claude Opus 5.5 completed 34 out of 50 security tasks, the same rate as having no safeguards at all. That’s the point. These tools are powerful, and Anthropic is trying to get them into the hands of defenders before attackers figure out their own workarounds.
High severity doesn’t automatically mean active exploitation
Here’s something worth understanding before you panic. A VulnCheck analysis published in late September 2026 looked at 300 vulnerabilities discovered through Anthropic or Project Glasswing. Of those, only 2 had been exploited in the wild. That’s 0.67%. The two confirmed cases were CVE-2026-26980, an SQL injection flaw in Ghost CMS, and CVE-2026-61500, a session forgery flaw in Rejetto HTTP File Server.
So AI is surfacing a lot of real vulnerabilities, but most of them aren’t being weaponized immediately. That doesn’t mean you can ignore them. It means you need a triage process, not just a patch queue. Critical and high severity items need attention first, and you need visibility into which of your systems are actually running the affected software.
The AI patch problem nobody is talking about enough
There’s a second issue buried in this story. Veracode found that roughly 44% of AI code generation tasks introduced a risky security vulnerability in tests. The average security pass rate across AI models sits at 56%, barely changed from 55% in their previous report. AI is finding bugs and also creating them at the same time.
If your developers or vendors are using AI to write or patch code, that code needs review. I’d also encourage you to read our earlier breakdown of why AI-generated patches carry their own risks for small businesses and what to check before you deploy them.
What to actually do right now
You don’t need access to Claude Opus to respond sensibly to this. A few concrete things help:
- Know what software you’re running. Ghost CMS and Rejetto HTTP File Server are niche, but similar flaws will hit mainstream tools. An accurate software inventory is step one.
- Prioritize critical and high severity patches on a defined schedule, not whenever someone gets around to it.
- If you use AI-assisted development tools or AI-generated code from vendors, ask about their security review process.
- Make sure your patch management covers third-party applications, not just Windows Updates.
Our post on preparing for larger and more frequent security patches covers the operational side of this in more detail if you want the specifics.
The volume of discovered vulnerabilities is only going to increase as more security teams get access to tools like these. Having a managed process in place now is a lot less painful than scrambling after something gets exploited. If you want a second set of eyes on your current patch posture, booking an IT assessment with Exine is a good starting point.
FAQ
How many vulnerabilities did Anthropic’s AI find in 2026?
Between April and July 2026, Project Glasswing identified at least 129,000 verified software vulnerabilities. An additional 5,500 were found through open-source scanning by October 2026. More than 33,000 of the total were rated critical or high severity. Anthropic believes the true count is likely five times higher.
Should small businesses worry about AI-discovered vulnerabilities?
Yes, but with context. Most discovered flaws aren’t immediately exploited. Of 300 Glasswing-related vulnerabilities analyzed by VulnCheck, only 2 saw active exploitation. The real concern is keeping your patch process fast enough to close gaps before attackers catch up to the same research.
Can AI-generated security patches cause new problems?
They can. Veracode testing found that roughly 44% of AI code generation tasks introduced a new security vulnerability. The average security pass rate across AI models was only 56%. Any AI-generated patch or code fix should go through a security review before it reaches production systems.
What is Anthropic’s Cyber Verification Program?
It’s a tiered access program announced on October 7, 2026 that lets vetted cybersecurity professionals use Claude AI models with reduced restrictions. The three tiers cover defensive work, authorized penetration testing, and specialized safety testing. It’s designed to give defenders access to the same AI capabilities attackers might misuse.
