TL;DR: The vuln surge (‘The Vulnpocalypse’) is real. The breach surge isn’t. YET. Vulns were never the bottleneck; the economics are.
Everybody is talking about the tsunami of vulnerabilities, triggered by the application of advanced AI models against our systems, that is about to hit our cyber shores. However, it seems that currently the surge in vulns does not translate to a surge in breaches.
Press enter or click to view image in full size
This is the cyber version of the Fermi paradox: if attacks are so easy to pull now, then where is everybody?
Let’s talk facts: Vulns and breaches in the numbers
We all “feel” that vulnerabilities are thriving, but breaches are not. However, we must work with concrete numbers and not with gut feelings.
It is easy to prove that the vuln surge (“The Vulnpocalypse”) is indeed real. Software vulnerabilities are tracked through assigned CVE (Common Vulnerabilities and Exposures) numbers and their risk is measured by the CVSS (Common Vulnerability Scoring System) score.
On April 7, 2026, Anthropic made Claude Mythos Preview available to select security researchers and critical infrastructure partners. Since then, the count of important CVEs (Critical and High severity) has surged several-fold.
Press enter or click to view image in full size
On the other side of the equation, measuring breaches in relatively real-time is a harder task. To do so I have utilized public companies SEC filings. The SEC’s cybersecurity disclosure rule took effect December 18, 2023. It forces every public company to publicly tell investors about a material cyber incident on Form 8-K within four business days.
Press enter or click to view image in full size
While there is an uplift in reported material breaches in Q2 and Q3 of 2026, it is dwarfed by vulns’ several-fold surge.
The materiality criterion serves another critical purpose: it reduces the noise of immaterial breaches and lets us focus only on breaches that have a business impact and should actually concern defense teams and the board.
Therefore, we can conclude that everybody’s gut feelings were correct: Vulns explode, but breaches do not.
But why?
Vulns are not the technical bottleneck
It’s important to understand that software vulnerabilities and even their exploitation are not the full story. A targeted cyber attack is a multi-step process. Mandiant captured this notion more than a decade ago, in their widely accepted “Targeted attack life cycle” model.
Press enter or click to view image in full size
This model clearly shows the multi-step nature of the targeted attack process, in which vulnerabilities can be relevant in the “Initial compromise” stage and maybe other stages. However, vulns are not even mandatory and can be replaced by stolen credentials, for example.
Having said that, even if the AI-driven Vulnpocalypse is not solving the targeted attack bottleneck, it seems that AI can solve all other technical steps within the “Targeted attack life cycle” model. And in fact, in one widely discussed demonstration, OpenAI agents compromised HuggingFace in an attack that walked through every stage of the model, from initial recon to complete mission.
So why isn’t it happening more in the real world? Are there some non-technical bottlenecks?
The financial bottleneck
The Verizon Business DBIR (Data Breach Investigations Report) report is one of the most widely used sources for breaches’ trends high-level analysis.
The 2026 DBIR report says that financial gain is by far the main motivation for cyber breaches.
Press enter or click to view image in full size
For the financially motivated attackers, the Complete mission step of Mandiant’s Targeted attack life cycle model is not the end of the story. Even once the attackers gained access to the victim org’s secrets, they still need to monetize that stolen information and launder the received money so they can actually use it.
Press enter or click to view image in full size
A great example of the importance of these financial value extraction steps is Ransomware. Ransomware elegantly addressed both of these issues in the previous decade. Monetization was solved by selling access to the encrypted data back to its rightful owners, and Money Laundering was solved by the use of a nascent cryptocurrency economy whose fund movements regulators were still baffled about how to monitor. In a way, ransomware is much more of a genius business model than a technical attack.
Therefore, the most effective way to fight it is by breaking this business model. According to chainalysis’ recent report, ransomware revenues are dropping, mainly due to government (regulators, Law enforcement) actions to limit attackers’ ability to monetize their attack and to successfully launder the received ransom:
Improved incident response and increased regulatory scrutiny have helped reduce payout frequency.
Effective international action against ransomware operators, infrastructure, and laundering networks has constrained some revenue flows.
Press enter or click to view image in full size
Clear evidence that Initial compromise techniques such as stolen credentials and vulnerabilities were not the bottleneck, can be found in their declining price in Initial Access Broker (IAB) markets, even before the offensive AI tools introduction, as stated in the same chainalysis report:
Cybercrime prevention firm Darkweb IQ told us that, from Q1 2023 to Q1 2026, the average price for victim access declined from approximately $1,427 to $439. This drop reflects a market characterized by more competitive pressure and automation as the core shaping force. “We are seeing industrialized access pipelines, AI-assisted tooling, and a proliferation of infostealer logs that lower the barrier to entry, which has resulted in an oversupply of cheap but operationally constrained inventory that floods the market and depresses pricing.
If the Initial compromise step was truly the bottleneck in the Targeted attack life cycle, the new surplus in supply should have been absorbed by the demand and not cause a price drop. New AI-found vulnerabilities and exploits just pile onto an initial-access surplus that is already going largely unused.
Therefore, I believe that these financial stages are the true bottleneck, that prevents the newly introduced offensive AI tools from translating into actual breaches.
Alternative explanation 1: Attackers lack access to the relevant AI tools
An alternative explanation for the lack of Breachpocalypse can be that it is merely a matter of time: attackers simply don’t have Mythos-class capabilities in hand yet, and once they do, the breaches will follow.
I don’t find this explanation plausible, as the capability has already reached attackers, either through open weights models or by bypassing guardrails:
- Bypassing guardrails: Where frontier access is required, attackers are already reaching it, by deception. Hacktron’s team showed this cleanly: Claude Opus 5 refused to write an exploit against a remote instance, so they proxied their own target through a URL dressed up as a CTF challenge, and the refusal evaporated. Anthropic’s own recent September 2026 threat report tells the same story from the other side, documenting state and criminal actors running autonomous vulnerability discovery and end-to-end intrusions, one escalating from a single stolen token to full cloud control in about three hours, all on standard, generally-available models.
- Open weights models: Mythos was indeed a watershed moment, but open-weight models typically close the gap with closed-weight models within about four months on average, according to epoch.ai.
Press enter or click to view image in full size
So the capability is already in the hands of attackers: They can discover, exploit, move laterally, and exfiltrate at machine speed, today, with tools they already possess. And yet the breach counts do not surge.
Alternative explanation 2: The Breachpocalypse is already here, but we don’t see it
The gut feeling that “breaches are not surging” is popular, and we used SEC filings to support it. However, it might be that breaches are already happening but have not been discovered yet, and thus not reported.
This “submerged breaches” hypothesis is, of course, a possibility that only time will tell, and we will know much more a few quarters from now — but by then it might be too late to act, and we must assess today based on the available data.
That said, I don’t find it highly plausible. A breach surge driven by financially motivated attackers would have to monetize eventually, and monetization is loud: extortion demands, leak-site postings, fraud losses, and account takeovers all surface within weeks, not quarters, and must be reported to SEC if material.
A word of caution
Naturally, my thesis that the economics are stopping the breach surge, even if supported by relevant evidence, might be wrong. But even if it is true it does not mean that AI is not currently changing the cybersecurity threatscape:
- There can be an increase of non-financially motivated attacks, as they are not limited by the aforementioned financial bottlenecks.
- For the Financial sector, monetization can be trivial with attackers directly stealing funds. This is especially true for the Cryptocurrency sector in which money laundering can be easier through the use of mixers and bridges to privacy preserving chains.
- New entrants to the cyber offensive market may miscalculate and put their focus on the new technical opportunities enabled by AI and neglect the financial related steps. This can cause an increase in the number of breaches until the monetization gap is fully realized.
- Even if AI is not changing the end result of breaches, it may change the internal balance of techniques within the “Targeted attack life cycle model” (e.g. vulns vs. stolen creds usage) and the velocity of this process. Therefore, defenders must re-assess and re-design their defenses accordingly.
Furthermore, the financial bottlenecks can be brittle and not controlled by potential victim. If attackers come up with a new way to monetize and launder their earnings, it will break this current “Cyber Dam” and we will see all of the attackers new capabilities deployed at once overflowing our defenses.
Summary and Key takeaways
We have shown that while AI is transforming the cyber threatscape,
- Vulnerabilities were never the bottleneck in targeted attacks.
- The real bottlenecks are economic.
Consequently, these are our recommendations to accommodate the AI-driven cyber future:
For governments:
Limiting cybercrime actors’ ability to monetize their attacks and to launder their earnings is a crucial part of protecting our digital lives. It’s more effective to fight mosquitos by drying their swamp habitat than trying to go after them individually.
For defenders:
AI is changing the threatscape and you should accommodate it. Attacks will be faster and may change the nature of the techniques used. Therefore, new concepts might be needed, and not just “faster horses” of doing the same just faster.
If you are lucky enough to not have non-financially motivated adversaries and your company is not in the financial sector you may have some extra time to prepare.
My advice to you: Make good use of this time; don’t sit on your hands. Winter is coming, even if it is currently withheld by the financial dam.