Skip to main content
Return to TrendAI 保安網誌
AI & emerging technologies Vulnerabilities and exploits

The 450% Problem: How AI Annihilated Your Patch Window

AI is accelerating vulnerability discovery and shrinking the window between patch and exploit. Here’s how to prepare your defenses before disclosure.

Perspectives AI LLMs Emerging technologies Exploits & Zero-Days

For years, the security community operated on a comfortable assumption: finding vulnerabilities was hard. It required rare expertise, months of research, and a deep familiarity with specific codebases. It’s one of the reasons TrendAI™ Zero Day Initiative™ (ZDI) financially compensates researchers who report bugs through our program.

That assumption is no longer safe. Artificial intelligence is fundamentally changing who can find bugs, how fast they can find them, and what defenders must do in response. The data we’re seeing at ZDI, along with the data coming out of Microsoft’s monthly Patch Tuesday, releases tells a story that should concern every security team on the planet.

Let me put it plainly: AI is collapsing the vulnerability discovery timeline, and the window between a patch’s release and active exploitation is shrinking with it. If you’re not patching faster, you’re losing ground. And as I’ll argue, patching faster is only the start.

The numbers don’t lie

Start with Patch Tuesday. For years, Microsoft’s monthly update cycle addressed a predictable range of CVEs. There were enough to keep administrators busy, but it was manageable. That changed. Recent Patch Tuesdays have routinely crossed into four-digit CVE counts, with several months setting records that would have seemed implausible five years ago.

The volume isn’t driven entirely by AI, but AI is unquestionably a catalyst. Automated vulnerability discovery tools, many of which now incorporate large language model reasoning, are helping researchers surface classes of bugs that previously required expert intuition to identify. More researchers finding more bugs faster means more CVEs. It’s a pipeline, and AI just widened it considerably.

The Pwn2Own™ competition tells the same story from a different angle. At Pwn2Own Berlin this spring, we saw a record number of entries. Close to 100 possible entries arrived in the final 72 hours before registration closed and, for the first time in the contest’s history, we had to turn people away. We also witnessed successful exploitation attempts across a wide range of targets, often fueled by the use of agentic AI in some form or fashion.

Five months later at the Pwn2Own competition, contestants disclosed 98 unique zero-days and earned more than $1.2 million, with AI infrastructure and AI coding agents now standing as contest categories alongside phones and printers.

Submissions to our program have grown 450% year over year with no sign of slowing. If anything, AI will accelerate the trend further. Pwn2Own contestants are increasingly leaning on AI-assisted analysis to accelerate their research and, notably, to produce the documentation required to register an entry. What used to take weeks of reverse engineering can now take days. The practical ceiling on how many targets a skilled researcher can meaningfully prepare for during a competition window has risen dramatically.

But there’s a nuance worth calling out: the most impactful work at Pwn2Own is still coming from advanced researchers finding bugs that AI-assisted code review never caught. AI is a force multiplier, not a replacement for deep human expertise. The bugs being demonstrated on stage are not the low-hanging fruit. They are the difficult, in-depth vulnerabilities that require a deep understanding of how systems behave under stress.

That’s a reminder that even organizations running AI-powered security tooling cannot assume their code is clean. It’s also why that human expertise is worth more than the bug count. Knowing how a class of systems breaks is what lets you predict where the next attack will land, and prepare for it before anyone files a CVE.

And then there’s what we see directly at TrendAI™ ZDI. More researchers are participating, and more bugs are being found in categories we’d consider well-trodden (e.g., memory corruption, use-after-free, type confusion) because AI tooling is helping researchers iterate through potential vulnerability patterns at scale. We’re also seeing reports from researchers who are newer to the field.

AI is democratizing vulnerability research. That has real upside for the ecosystem in the long run. In the near term, it means more bugs are entering the disclosure pipeline than the industry has historically been equipped to process.

What AI is actually doing

To understand the urgency, it helps to understand the mechanism. Modern AI tools, whether purpose-built security tools or general-purpose code reasoning models, are particularly effective at a few tasks that were previously bottlenecks in vulnerability research.

First, they accelerate variant analysis. Once a class of bug is understood, AI can scan for similar patterns across large codebases far faster than any human. A single disclosed vulnerability routinely spawns AI-assisted searches that uncover multiple related issues in the same component.

Second, they lower the barrier to reaching new attack surfaces that were harder to audit in the past, such as embedded systems and complex legacy codebases.

Third, the same tools available to researchers are available to threat actors. The asymmetry that historically favored defenders (finding bugs is hard; patching is at least structured) is eroding. Nation-state actors and well-resourced criminal groups are deploying AI aggressively.

Notice what every one of these dynamics has in common: the defender’s clock doesn’t start until disclosure. Discovery, patch release, triage, testing, deployment. Each handoff was tolerable when finding bugs was slow. But now AI is putting pressure on every one of them simultaneously, making the sequence itself the new vulnerability.

What you need to do now

If the sequence is the problem, the answer isn’t to run it faster. It’s to start before it begins. The organizations that survive this shift will be the ones that treat AI-driven vulnerability discovery as a present operational reality and do most of their defensive work before a CVE ever has a number.

First, know your attack surface before the attacker does. AI-assisted attackers will find your unpatched systems faster than you expect. You need to know your own environment at least as well as they do, and that environment now includes the AI infrastructure and coding agents researchers are already exploiting on stage. Visibility isn’t a hygiene exercise anymore. It’s the foundation for everything below, because you can’t anticipate an attack path through systems you can’t see.

Second, prioritize by what could actually succeed in your environment, not by severity score. Not every CVE in a bumper Patch Tuesday crop carries equal risk, and a 9.8 in a component you’ve already segmented off may matter less than a 7.5 sitting on a path to your crown jewels. The question isn’t “how bad is this bug?” but “could this attack path work here, and what would it reach?” Use threat intelligence to separate the emergencies from the noise. TrendAI™ ZDI’s vulnerability data and CVSS supplemental scoring exist precisely for this purpose, and TrendAI Vision One™ Cyber Risk Exposure Management lets you test that question against a model of your own environment rather than guessing.

Third, have the defense path ready before the patch exists. A virtual patch is a security rule you apply on the network or endpoint that blocks attempts to exploit a vulnerability without changing the vulnerable software. TrendAI™’s network-level virtual patching can protect you up to 115 days before a vendor patch is released. But the point isn’t just buying time after disclosure. It’s that the defense path can be mapped and staged in advance so when the advisory lands, you can immediately activate a response you’ve already tested.

Fourth, compress your patch cycles. Preparation doesn’t replace patching; it makes patching the last step instead of the only one. For critical and actively exploited vulnerabilities, which Patch Tuesday flags explicitly, that window needs to shrink to days, not weeks.

Fifth, scale through automation. CVE counts will keep rising, driven by AI in both the discovery and reporting phase. Even vendors who manage to keep pace with disclosure will have enterprise customers and end users who cannot. The math doesn’t work if you’re still treating patch deployment as a manual process. We learned this ourselves: ZDI now uses agentic AI to triage incoming submissions, so an analyst can keep up the pace. Routine updates need to be automated so that human attention can focus on reducing real business risk.

Sixth, learn from every cycle. Attackers adapt. Your environment changes. Every Patch Tuesday, every ZDI disclosure, every Pwn2Own result should feed the next round of preparation. Which control held? Which attack path did you miss? Where did the response take longer than it should have? Programs that treat each disclosure as a one-off emergency will keep getting surprised, whereas programs that treat each one as a rehearsal will get faster.

What’s changing at TrendAI™

This tectonic shift has changed how we think about defense at TrendAI™. For 20 years, the goal of TrendAI™ ZDI was to find the bug first and buy defenders time. That still matters, but it’s no longer enough. The starting point has to move earlier, to anticipating plausible attacks and having a tested defense path ready before disclosure happens. We call it Minus One: predictive cyber defense. Defense prepared, before zero day.

Minus One will be delivered through TrendAI Vision One™, bringing together the ZDI intelligence, customer environment context, and protection controls that make this kind of preparation possible. Read more about Minus One here

The vulnerability landscape grows more complex each day. AI guarantees that. But the organizations that will weather this shift are those that have prepared their defenses before disclosure, automated the routine, and freed their teams to act decisively on the critical. The era of slow is over, and the security programs still operating on that assumption are the ones threat actors are counting on.