AI Vulnerability Finding

A comment:

3% of ‘critical’ software security bugs, in all of recorded history, have been found in the last month. This is on pace to find another ~100,000 bugs this year alone, adding ~26% more vulnerabilities to the all time total Do you realize how fucking insane that is?

Yes. It is a new world out there.

Posted on November 2, 2026 at 7:02 AM • 17 Comments

Comments

Rontea • November 2, 2026 8:19 AM

It’s a reminder: the attack surface isn’t shrinking, and defenders need to be proactive, not reactive.

Nasrin • November 2, 2026 9:27 AM

Yes, this is insane. And of course the gap between detection and defense is becoming dangerously wide…

uh, Mike • November 2, 2026 9:45 AM

A decade or two ago at SANS, a presenter said that someday we would solve network security, because we have to.

(Sorry, I can’t remember who said it.)

Tyrone McWalluce • November 2, 2026 10:21 AM

This is why NSA wants to regulate AI like Mythos: Because it helps close NSA’s backdoors.

Jeff • November 2, 2026 10:28 AM

So many people are trying to act like this is a bad thing, it’s a good thing! More bugs aren’t necessarily being found because there’s more of them being created than before, the more likely explanation is that the bugs were already there and now we’re closing them at the fastest pace in history!

“This is on pace to find another ~100,000 bugs this year alone, adding ~26% more vulnerabilities to the all time total”

Or…perhaps it’s closing 100,000 more vulnerabilities than every before!

Clive Robinson • November 2, 2026 11:45 AM

@ Jeff, ALL,

With regards,

“More bugs aren’t necessarily being found because there’s more of them being created than before, the more likely explanation is that the bugs were already there and now we’re closing them at the fastest pace in history!
“

Actually both points are true…

Yes we are finding “new instances of bugs from known classes of bugs” at a much higher rate.

But that does not mean we are finding “new classes of bugs” (ie black swans).

But also “vibe coding” by AI is certainly generating “new instances of bugs” at a more alarming rate.

It’s why quite a few employers are now seeing “laying off humans with AI” as being not just horrifically unprofitable, but extremely dangerous to the business.

Yes, AI has niche advantages, but those niches are invariably not general in nature.

Even China is begining to see the dangers of even low cost AI.

None of this was unpredictable and quite a few have called it that way. However when you’ve handed over everything for a few tulip bulbs you had better be sure not only are they going to grow but that people are going to by the result sufficient to pay a “Return on Investment”(ROI).

Right now the ROI in “General AI” is not a certainty or even a likely hood based on the current economy…

Then there is “clearing up the mess of harms” the rush for AI has created… How many billions if not trillions that is going to cost companies that are already by normal forecasting methods “bankrupt”…

Rick • November 2, 2026 12:28 PM

I do wonder how many of those bugs found by AI are “real”…

I don’t think it changes the message or the overall idea. Theye are finding a lot of real stuff. But there’s also a lot of “inflated” findings when you start going through everything, a lot of not actually a problem, etc etc.

Just wondering if this is another metric-turned-into-target that will loose its meaning soon.

tfb • November 2, 2026 12:34 PM

If machine-generated (‘vibe coded’) code is introducing bugs then, conveniently, we now have machines which seem to be extremely good at finding them. We could even take a really radical step and make the code-producing machine and the bug-finding machine be the same machine.

freedom • November 2, 2026 3:03 PM

we now have machines which seem to be extremely good at finding them.

No, “we” don’t. Most of these claims are propaganda, marketing and outright fraud.

Also, one of the reasons, if not the biggest reason why the alleged bugs are found is because ridiculous amounts of computing power are being used. So this is just “brute forcing” a problem. Doesn’t seem like a very “intelligent” approach.

Clive Robinson • November 2, 2026 5:03 PM

@ Bruce, ALL,

You might want to read this from Robert Haas

Postgres 19 is late, and it’s AI’s fault

Over the past decade, the timing of PostgreSQL major releases has become very predictable: no earlier than the middle of September, and no later than the middle of October. This year is going to be different. PostgreSQL 19 most likely won’t be released until November, and it’s not impossible that the project could decide to postpone the release further.

“So, what happened?

The short answer is that LLMs happened. All software contains bugs, and PostgreSQL’s code base is large, complex, and old. Our git history – converted from CVS – now reaches back more than 30 years, and the first code was written a decade before that. There’s a huge amount of code, and that code contains a huge number of bugs. That’s always been true, and it hasn’t stopped PostgreSQL from being popular or successful. What’s different now is that it’s become vastly easier to find those bugs. Fixing them has also gotten easier, but not to the same degree.

Between April 8th, 2025 and September 25, 2025, there were 1037 commits to PostgreSQL’s master branch and 724 commits to REL_18_STABLE. In 2026, during the same date range, there were 1443 commits to the master branch and 1169 commits to REL_19_STABLE. In other words, if you look at the number of commits that went into the master branch only between feature freeze and September 25th, it actually declined slightly from last year to this year, but the number of back-patched commits jumped by more than 60%.
…
“

https://rhaas.blogspot.com/2026/10/postgres-19-is-late-and-its-ais-fault.html

Ismar • November 3, 2026 4:15 AM

I feel that same of us here keep falling into an old trap of overreacting to a most perceived immediate challenge in their lives, such as is AI at the moment. This results in rejecting it outright instead of evaluating its pros and cons following established scientific methods.It would be much more useful to identify how to use this new technology in a most beneficial way for purposes of ensuring our future is more sustainable and just for most of us and not only for a select 0.01% . And, yes I think it can be used for that as well

freedom • November 3, 2026 2:18 PM

This results in rejecting it outright

What we are getting is outright pro “AI” propaganda so i see no problem with completely rejecting it.

You wnat the critics to be more “fair” and balanced but you seem to be overlooking how blatantly biased the pushers are.

David • November 5, 2026 1:44 PM

3% of ‘critical’ software security bugs, in all of recorded history, have been found in the last month.

That’s the subset of bugs that were found and disclosed. How many were found and saved for future use?

Clive Robinson • November 5, 2026 4:58 PM

@ David, Bruce, freedom, ALL,

With regards,

“That’s the subset of bugs that were found and disclosed. How many were found and saved for future use?”

Also more importantly you’ve missed,

“How many were not recognised and passed over?”

As @freedom noted above,

“So this is just “brute forcing” a problem. Doesn’t seem like a very “intelligent” approach.”

It’s not an “intelligent approach” for a number of reasons but the important one is,

“It can only find what is ‘known’ or ‘small variations’ there of”

I’ve described this in detail in the past on this blog, long before AI got into the bug hunting game.

There are,

1, Known Classes of Attack.
2, Known Instances of Classes of Attack.

Which implies there are,

3, Unknown Classes of Attack.
4, Unknown Instances of Classes of Attack.

However a “Class” of Attack is a collection of “Instances” of Attack, that have “common characteristics”. So likewise one or more Classes can have common elements thus all vulnerabilities form a taxonomy. Which importantly is always going to be “incomplete”. The only real question –which we can not answer– is,

“How incomplete is our knowledge?”

One way of finding “New/missing Instances” is by probabalistic methods we call “fuzzing” (a term from before AI bug finding). In effect fuzzing and similar stochastic methods find instances that are “close in” to existing instances.

That is fuzzing is seen another way, is not even,

“Finding new ways to turn a screw, and most certainly not new ways to fasten things.”

Also take a moment to think over the difference between a “nail and a screw” and what the different types of nails and screws can have in common.

For instance a nail uses a “friction hold” that means it’s shaft can be either uniform in diameter or tapered in some way.

Screws were untill fairly recently also all tapered, because those that were uniform were called “bolts”.

The advent of fixing differing forms of a materials such as floor boards to joists gave rise to nails with non uniform tapered shafts.

Likewise the advent of fixing different materials such as sheet metal to wooden frames gave rise to both different screws as well as different nails. It’s why you have screws that actually have a “drill” as the first part of the screw and differing threading along the shaft of the screw.

So… you have the “Class of Screws” but with “sub-classes” that have varying type or “instances” of shaft / head.

Now think about other fastenings such as “rivets” and “anchors” for sheet metal and sheet rock/plaster respectively. Then threaded holes such as captured nuts for bolts that are put in sheet metal in little cages or like half rivets.

You can see how classes of instances form trees of sub-classes, but also how small changes to a known class does not cover major changes.

The same applies to attacks.

Whilst “fuzzing” gets you “minor changes”, it rarely gets you “major changes”.

Thus the issue is todays AI finds,

“Known classes and some minor variations”

It does not find,

“Unknown classes that are major variations”[1]

Thus Current AI systems are almost certainly missing a great many vulnerabilities even though we are finding “known vulnerabilities” at a prodigious rate.

It still takes a human mind to

“Invent or recognise the truly new”

Something that Current AI based on LLM and ML systems just can not do. Nor if you understand how they work fundamentally would you expect them too.

[1] These have been called “Black Swans” –in long past times– but reimagined by Nicholas Taleb, in 2001 or more descriptively “Unknown, unknowns” by United States Secretary of Defense Donald Rumsfeld back in Feb 2002.

The thing is that Current AI Systems have “no agency” thus actually “no world view” they are like the “Three blindmen and an Elephant”. That is “They don’t see, they feel parts out of context” thus describe it as being like a rope, a tree trunk, or a wall.

So if there is no “known thing” in an AI’s database / encoded in it’s weights, that is like the “Unknown thing” that comes into an AI’s sensing, then it can not match it thus misses it entirely or potentially badly misclassifies it.

This misclassification is often caused by the way the ML statistically forms “tokens” from examples fed into it via initial training data.

Wales are not clouds, but if all the pictures you classify as wales during training have clouds in them, how do you expect the ML process “to know” what is a whale and what is a cloud?

Jimmy • November 11, 2026 11:03 PM

The 3% statistic reframes the whole vulnerability-management debate: if a quarter of all critical bugs ever recorded appeared in one month, then reactive patching cycles are structurally behind, not operationally sloppy. AI finding bugs at machine speed while remediation stays human-paced means the detection-defense gap becomes the primary risk surface. The proactive point in the post deserves the emphasis – triage policies and patch prioritization written for weekly discovery rates simply do not survive daily five-figure discovery rates.

Leave a comment

Blog moderation policy

Login

Allowed HTML <a href="URL"> • <em> <cite> <i> • <strong> <b> • <sub> <sup> • <ul> <ol> <li> • <blockquote> <pre> Markdown Extra syntax via https://michelf.ca/projects/php-markdown/extra/

Sidebar photo of Bruce Schneier by Joe MacInnis.