In May 1998, seven hackers from L0pht told a Senate committee they could take down the Internet in about 30 minutes. While that may seem far-fetched in 2026, given the state of cybersecurity at the time, no one in the cybersecurity community in 1998 was surprised that this sort of vulnerability existed.
If you were involved in either offensive or defensive security in any capacity back in the 90s or 2000s, you probably remember the cadence. Someone found a devastating vulnerability in software that everybody depended on. You dropped what you were doing, figured out where you were exposed, and either raced to fix it (if you were in the early cybersecurity industry of the time) or exploit it (if you were in EFNet IRC channels). Then it happened again the next week.
In 2002 alone, serious remote vulnerabilities hit some of the most important software powering the Internet: Apache, OpenSSH, OpenSSL, and BIND. Web servers, secure connections, DNS: pick a foundational piece of the Internet, and chances were somebody was having a very bad week because of it.
In conversations with security teams the past several months, it’s clear that this feeling is coming roaring back. The kind of vulnerability that used to clear a security team’s calendar for an emergency response is starting to feel less like an exceptional event every several years and more like how teams have spent the majority of 2026.
Opus 4.5/4.6 marked the beginning of a radical change in the economics of offense. Work that required scarce expertise and substantial time is becoming orders of magnitude faster, cheaper, and easier to repeat, and the results are showing up in key software packages that have been audited to death over the past couple decades (or so we thought).
There’s historically been a running joke among security teams that zero-days get all the attention, but the 550-days are what actually cause most of the breaches. That joke has had an impressively long shelf life. But we have unfortunately re-entered a period of time where we have to worry about vulnerabilities at both ends of the time spectrum, both zero-days and 550-days.
Luckily, we’ve been through this before. Taking the experiences from the first prolific zero-day era and applying it to today, combined with recent conversations with the CISOs of the most AI native security programs, a few themes are emerging about what the next few years will look like:
1. Worms 2: They’re back!
The early 2000s saw a wave of “Internet worms” which were self-propagating malicious programs that would find a host with a known vulnerability, compromise it, and then use the newly infected host to start scanning for more vulnerable systems to replicate itself to. In the earliest versions of these sometimes they generated so much network traffic due to so many infected hosts looking for new vulnerable ones that they would knock entire networks offline. We largely saw this style of attack drop off in the late 2000s as key parts of the Microsoft ecosystem got more secure and stopped having nearly as many remotely exploitable vulnerabilities.
If Internet worms coming back won’t make it feel like the 90s and 2000s again, nothing will. As AI wildly decreases vulnerability discovery and exploit development time, but patching timelines haven’t caught up yet, we will see windows of opportunity for Internet worms to make a return. Unlike SQLSlammer, Sasser, and others from the early 2000s era though where the goal was often simply self-propagation, this time we’ll see more commercially minded goals of the attackers stealing API keys for inference which is rapidly displacing crypto mining as the action of choice for attackers when using stolen resources, ransomware, etc.
2. It’s the Windows XP patching problem all over again
In the early 2000s, an unpatched Windows machine connected directly to the Internet would get compromised before it finished downloading the updates it needed to block the attacks. At the time it was such a problem that SANS even published a guide called Windows XP: Surviving the First Day.
The AI era version of this problem won’t just apply to operating systems, it’s already applying across the entire software ecosystem. Up until now, enterprises have held off on automatic updates of most software packages for two reasons. First, because the risk of it breaking things has outweighed the security risk of not updating. AI unlocking offense at machine speed will change the calculus here and we’ll see enterprises increasingly apply patches at orders of magnitude quicker speeds. Secondly, because it’s a much harder problem than these few sentences make it sound. The sheer amount of software inside an enterprise is staggering, combined with the fact that a large percentage of it is often no longer actively maintained. To that end, we are excited to see new startups in this space aimed at helping enterprises patch orders of magnitude faster and more effectively.
3. Startups will develop around hardened infrastructure components
In the prior version of this sort of security environment, we saw multiple open source projects and startups form around the idea of highly hardened infrastructure components. For example, OpenBSD built its identity around correctness, security auditing, and secure defaults and for quite a while was held up as the gold standard in secure operating systems (I ran it as my desktop OS in the early 2000s, which tells you everything you need to know about how popular I was in high school.)
We will likely expect a new generation of products whose central promise is taking a critical piece of infrastructure, and delivering a highly hardened version of it for enterprise use. The previous generations of these had a hard time commercially, but we will see if the changes in the security landscape unlocks a significantly larger market this time around.
High level thematic predictions are fine and all, but for the security teams in the midst of dealing with this right now, where should they start?
1. Aggressively test yourselves. Now.
Not an annual pentest for compliance purposes. In fact, most security teams we talk to are reorienting their pentest efforts to be as continuous as possible, and reclassifying their legacy pentest services as an annual compliance exercise.
For the era we are in, security leaders need to embrace no scope, full red team testing using the latest tools that attackers are actively having success with at this moment. Legacy approaches to vulnerability management, scanning, pentesting, etc. are simply not discovering the issues that we’re seeing lead to the daily breaches of the moment. Instead, attackers are leveraging tools like Hinoki Security to get access to specialized post-trained security models that allow specifically for modern red team testing without tripping over model refusals.
In addition to external testing, if there was ever a time to also do aggressive internal testing as well it’s now due to the wave of zero days coming that will make it easier and easier for external attackers to gain an initial foothold inside an environment.
2. Defense in depth shifts to overlapping detection and speed
“Defense in depth” has been a common mantra for as long as there has been a security industry, but what it has meant on the ground has changed over time. When we as an industry first started rolling out security controls it often meant simply that in addition to deploying a firewall, also deploying an IDS behind it to try to detect attacks that slipped through. As corporate environments grew and grew over the last 20+ years, the reality that many CISOs privately share is that it’s often only during a breach that you realize that one or two layers of your controls weren’t functioning correctly, weren’t tuned, or in plenty of cases were simply accidentally switched off entirely. As a result, over this timeframe the practical point of defense in depth often became about layering enough controls where even if one or two didn’t function as intended there was still a decent chance that one of your controls would perform correctly and either help you prevent, detect, or respond to an incident. As you might imagine, all of this functioned quite slowly, but even though attackers tended to function quicker it was still roughly at the same order of magnitude of speed so there wasn’t a wide imbalance.
With AI extensively assisting all phases of the offensive capabilities now, we’re shifting into a new goal of defense in depth: one where, even if all your layers of controls are functioning correctly, the use of zero-days, new techniques, and outright speed will likely bypass one or more of your fully functioning controls. In this new world, security teams need to get into a mindset of not only deploying overlapping controls to raise the chance of detection, but also optimizing their detection and response capabilities around an AI native level of speed that we’ve simply never had to function at before.
3. Reboot your security stack for the AI operating environment
The shift to cloud was the last time defenders needed to reboot their security stack from the ground up, and in some ways the need was much more obvious vs. today. As a CISO in that shift, you had previously been buying physical appliances for many of your security controls. Suddenly there was no datacenter rack to plug them in to anymore, so you started looking for cloud native security controls that were entirely software.
The AI era changed virtually every part of the enterprise process and technology usage, and an equal level of reboot is needed for many of the most prominent security controls across organizations. Similarly to how we went from generation one of malware in the anti-virus era, to generation two in the behavioral based EDR era, we’ve now entered the third generation of malware and endpoint security which is the agentic malware era. On the identity side, we no longer need to simply authenticate and authorize human actions but instead provide controls for how agents function inside the enterprise. In other areas like open source software supply chain, we’ve seen the number of attacks absolutely explode due to the meteoric rise of agentic coding pulling in an increasing number of open source packages regardless of risk.
Just like shifting from physical appliances to cloud based architectures, all of these areas require a fundamentally new AI native architecture. As a result, they’re being tackled by new companies who can build for this era from the ground up, folks like Neo for endpoint-native AI security, Keycard acting as the new Okta for agents, and Socket protecting organizations from the sudden explosion of open source supply chain attacks.
Blunt truth be told, virtually every security leader we talk to on offense and defense expects the next several years to be a deluge of new vulnerabilities being published. Simultaneously, the time delta from discovery, to working exploit, to active attack campaigns will continue to collapse resulting in more and more breaches in the headlines.
While it may sound like unprecedented circumstances, the reality is there’s plenty we can learn from when the security landscape was like this 20+ years ago. Much like previously, it affords us the opportunity to drastically push forward the security capabilities of defenders by leveraging new AI native approaches across the entire stack. It will take new companies, new architectures, and a lot of work on the software we already depend on. But coming out the other side we’re going to have unlocked defensive capabilities that were only a pipe dream previously, which makes this quite possibly the most extraordinary time to be building in cybersecurity. Onward!
Special thanks to Joel de la Garza for all his help on this piece.