The Vulnerabilities We Can See Aren’t the Ones That Worry Me

AI is making serious software vulnerabilities easier to discover than ever. The public CVE explosion may be just the tip of an iceberg, and we have no idea what's happening beneath the surface.

Earlier this week, two critical vulnerabilities in PHP made headlines across the security community. The flaws, collectively nicknamed PHP2Shell, had been hiding in one of the most widely used programming languages for years. Researchers demonstrated how they could potentially allow attackers to execute code remotely on affected servers without authentication.

But one of the most remarkable details wasn't the vulnerability itself. It was how easily an AI agent could be directed to find it.

Dan Lorenc, CEO of Chainguard, described independently rediscovering the vulnerability after hearing that something significant was coming in PHP. His instruction to the AI agent was remarkably simple:

"let's look for pre-auth RCEs in PHP."

That's essentially it. Go find serious vulnerabilities in PHP. No elaborate prompt engineering, no detailed explanation of PHP's internals, and no guidance pointing to the actual flaw. The underlying technology was sophisticated, of course, but the effort required to direct it was astonishingly small.

Then, just days later, another major vulnerability emerged. This time it wasn't PHP.

Researchers disclosed CVE-2026–111497, a critical vulnerability in Brotli, Google's widely used compression library. Initially drawing attention because of its impact on Node.js, the issue has since been traced across a much broader ecosystem, including technologies such as Bun, Envoy, Istio, curl, and Python packages. Aikido's latest analysis describes how the potential impact extends well beyond a single application or framework.

And once again, AI was involved in discovering it. This time, OpenAI's Codex Security team identified the flaw while its agents were examining software in OpenAI's own environment.

Within a matter of days, AI-assisted research helped expose serious vulnerabilities in two foundational pieces of software used throughout the internet. These weren't obscure projects or newly introduced bugs. They were weaknesses hiding in mature software that had been examined by security researchers for years.

I spent the early part of my career working in information security and application security, and I find the pace of these discoveries extraordinary. Finding serious vulnerabilities in mature software once required substantial expertise and often weeks or months of investigation. Now we're watching AI systems uncover them at a pace that would have been difficult to imagine just a few years ago, sometimes with little more than a one-sentence instruction from the person directing them.

And that brings me to what I find genuinely frightening about all of this.

If this is what researchers are finding with relatively little effort, imagine what sophisticated attackers are doing with dedicated infrastructure, experienced security teams, and months to work on their targets. What are well-funded nation-state operations discovering? How many automated agents might already be searching through widely deployed software, quietly identifying weaknesses that nobody else knows about?

And how many of those discoveries have already been used to compromise systems without anyone realizing it?

We don't know. That's the problem.

These are the vulnerabilities we're hearing about because someone chose to disclose them. What about the ones being found by people who have absolutely no intention of telling us?

I think we're watching the tip of an iceberg emerge, and we have no idea how much is underneath it.

The vulnerability discovery curve has changed

In my previous post about AI-powered computer worms, I wrote about the possibility of malware that could discover vulnerabilities on its own as it spreads, rather than relying entirely on exploits its creator already knew about. At the time, much of that concern was about where things might be headed.

The vulnerability discovery part of that equation is already happening, and the public numbers are staggering.

I pulled data from the official CVE list and compared weekly disclosures across 2024, 2025, and 2026. For most of 2024 and 2025, the pattern was relatively consistent, with a few hundred to around a thousand new vulnerabilities published each week. Then, around July of this year, the curve changed dramatically.

Weekly CVEs published to the official CVE list. The 2026 series runs through the week of October 12–18.

In 2023, there were 28,847 published CVEs. That grew to 39,924 in 2024 and 48,151 in 2025. Through October 24 of this year, we're already at 86,403, roughly 80% more than all of last year, with over two months remaining.

The more recent acceleration is even more dramatic. Over the latest 12 weeks, 38,639 CVEs were published, compared with 10,937 during the corresponding period last year. That's approximately 3.5 times as many. The week of October 12 alone produced 5,460 disclosures, the highest weekly total in this dataset.

Now, more published CVEs don't necessarily mean software has suddenly become less secure. Changes in reporting practices, assignment processes, and disclosure backlogs can all influence these numbers, and we can't attribute the entire increase to AI.

But the timing is difficult to ignore. We're seeing this extraordinary acceleration at precisely the moment AI models and autonomous agents are becoming remarkably capable of reading code, identifying vulnerabilities, generating tests, and helping researchers develop working exploits.

Microsoft's frontier AI vulnerability research has helped uncover vulnerabilities associated with 140 Windows CVEs between May and September. Anthropic, OpenAI, and other organizations are producing similarly impressive results against widely used software.

These aren't theoretical bugs. We're watching AI systems uncover genuine security flaws in software that has been around for years, sometimes decades.

Which brings me back to the iceberg.

The market beneath the surface

There's another side to vulnerability research that doesn't get nearly as much public attention. Undisclosed vulnerabilities, particularly those accompanied by reliable exploits, are valuable commodities. There's an established marketplace where researchers, brokers, security contractors, governments, and criminal organizations buy, sell, and trade access to weaknesses that software vendors don't yet know exist.

This isn't some obscure corner of the dark web. Parts of this industry operate as legitimate businesses, with professional researchers, government contracts, and substantial acquisition budgets. Other parts operate in the gray and black markets, where buyers are considerably less transparent about their intentions. Companies like Zerodium and Crowdfense have publicly demonstrated how much buyers are willing to pay for high-value exploits, sometimes reaching millions of dollars for sophisticated capabilities.

The economics are fairly straightforward. An undisclosed vulnerability can provide access to systems that everyone else believes are secure. Its value depends on factors like the popularity of the affected software, the reliability of the exploit, how much access it provides, and whether anyone else knows about it. A complete exploit chain capable of compromising a modern mobile device, for example, is a very different commodity from an isolated software bug that might be difficult to exploit.

And in many cases, secrecy is central to maintaining that value. Once a vulnerability is publicly disclosed and patched, the window for exploiting it as a zero-day begins to close. An exclusive capability can be worth far more than one that has been widely circulated or independently discovered.

This market existed long before AI, and it has been studied for years. A RAND study of privately held zero-day vulnerabilities found that exploits in its historical dataset had an average lifetime of approximately 6.9 years before the underlying vulnerabilities were publicly disclosed or patched. That doesn't establish how long vulnerabilities remain secret today, but it demonstrates how valuable prolonged exclusivity can be.

Now introduce AI into that market. If an AI agent can independently rediscover a critical PHP vulnerability with a one-sentence instruction, imagine what professional vulnerability researchers can accomplish with specialized models, dedicated infrastructure, and continuous automated analysis. For organizations already in the business of discovering and monetizing exploits, AI has the potential to dramatically improve the economics of their operations.

And unlike responsible disclosure programs, the objective in parts of this market isn't to publish a CVE or help a maintainer release a patch. It's to develop a capability, preserve its usefulness, and potentially monetize it.

This is why I find the public CVE explosion so unsettling. We're watching one side of the vulnerability discovery ecosystem accelerate dramatically, while having very little visibility into a private market that has operated for decades and has powerful financial incentives to remain invisible. What happens to the scale of that market when the cost of discovering vulnerabilities begins to collapse?

And discovery is only part of what we can't see. A privately held exploit doesn't necessarily remain unused. It may be deployed selectively against valuable targets, incorporated into broader attack campaigns, or used to establish persistent access that remains undetected for months or years. We already know that sophisticated intrusions can remain undiscovered for extended periods. What we don't know is how AI-assisted vulnerability research might be changing the scale and sophistication of those operations.

Google's Project Zero has been raising this problem for years. In its research on zero-day exploits used in the wild, the team has highlighted a fundamental unknown in information security: how many zero-day attacks are actually detected? Its analysis of the detection deficit explains why publicly documented zero-day exploitation provides an incomplete picture of what's really happening.

That research predates the current advances in AI. If we already struggled to understand the true scale of zero-day exploitation when discovering these vulnerabilities required substantial expertise and resources, what happens when AI dramatically lowers those barriers? We may be watching public disclosures accelerate while having even less visibility into how quickly offensive capabilities are advancing behind the scenes.

If public vulnerability disclosures are suddenly growing at three or four times the previous rate, what might be happening inside an established industry whose financial incentives depend on keeping its discoveries secret? And perhaps more importantly, how much of that activity might already be happening without anyone realizing it?

Even the good guys are struggling to keep up

Another revealing example comes from Anthropic's vulnerability research.

According to its coordinated vulnerability disclosure dashboard, Anthropic had reported 6,157 vulnerabilities across 591 open-source projects as of October 2. Of those, 516 had been patched upstream, while 219 CVEs and 365 GitHub Security Advisories had been issued.

Those numbers aren't directly interchangeable. Not every finding warrants a CVE, some advisories overlap, and parts of the disclosure process are intentionally delayed to give maintainers time to address the issues. But the gap between the volume of findings and the number of completed fixes is striking.

What's striking here isn't just how many vulnerabilities Anthropic is finding, but how difficult it is for the software ecosystem to keep up. Even when discoveries are reported responsibly, fixes require engineering work, testing, coordination, and deployment. AI is accelerating discovery far faster than many of those processes can realistically change.

Now consider organizations that have no interest in disclosure. They're not waiting for maintainers to validate findings or release patches. Their incentive is to turn discoveries into reliable capabilities while preserving their secrecy. And unlike defenders, who must secure entire systems and their dependencies, an attacker often needs only one exploitable weakness.

What happens when discovery moves at machine speed?

This brings me back to the concern I raised in my earlier article about AI-powered worms.

Traditional computer worms generally relied on vulnerabilities and exploit techniques that their creators already knew about. Imagine a sufficiently capable autonomous system that could inspect the software it encounters, discover previously unknown vulnerabilities, and develop ways to exploit them as it moves between targets.

We're not necessarily at that point today. Finding a vulnerability and reliably exploiting it are different challenges, and both can require substantial technical work. But the PHP2Shell research demonstrates how AI is beginning to reduce some of those barriers, even for complex vulnerabilities that have existed for years.

Historically, the difficulty of finding certain vulnerabilities provided a degree of incidental protection. It was never something security professionals should have relied on, but it was a practical reality. Some bugs survived for years because finding them required more expertise, time, and effort than most people could provide.

That protection is getting weaker.

And this is where I think the security industry needs to move faster. We should be using the same AI capabilities to continuously examine our own software, identify weaknesses, validate findings, and accelerate patch development. We also need to assume that important systems may contain vulnerabilities somebody else has already discovered, even when there's no public advisory.

Traditional security fundamentals like isolation, least privilege, defense in depth, and continuous monitoring become even more important when we can no longer assume that undiscovered vulnerabilities will remain undiscovered for very long.

What we can see isn't what worries me

I'm genuinely excited about what AI is making possible in security research. Watching machines uncover vulnerabilities that have been hiding in software for years is remarkable, and every responsibly disclosed finding gives us another opportunity to make the systems we depend on safer.

But having spent the early part of my career in security, I can't look at these numbers without thinking about what they don't tell us.

We're seeing an unprecedented acceleration in publicly disclosed vulnerabilities, while the expertise and effort required to discover serious flaws are falling. What used to demand highly specialized research can increasingly be initiated with something as simple as a one-sentence instruction to an AI agent.

Now imagine what the people who really know what they're doing are accomplishing with these same capabilities, more sophisticated tooling, and virtually unlimited motivation to keep their discoveries secret.

We don't know how many vulnerabilities they're finding, how many they're stockpiling, or how many may already have been used to compromise systems that nobody realizes are compromised.

And that's the part that worries me.

The CVE explosion may not be the full picture of what's happening in vulnerability discovery. It may simply be the first time we've had a good look at the tip of the iceberg.

The vulnerabilities being publicly disclosed aren't necessarily the ones that keep me up at night.

It's the ones being found right now, quietly, by people who have absolutely no intention of telling anyone.

Data note: CVE statistics are based on the author's analysis of the official CVE List through October 24, 2026. Published CVEs, assigned identifiers, reported security findings, and exploitable vulnerabilities are distinct measures.