What AI Actually Did to Cybersecurity in 2026

Do you remember the cyber-apocalypse? The one that powerful AI models were supposed to cause. AI was going to write unstoppable malware, break every sandbox, discover all of the zero-days, and make cybersecurity professionals obsolete. Well, it’s 2026, and there’s no cyber-apocalypse, at least not yet. But things did change, and I want to walk through how they changed.

Deepfakes actually fooled a bunch of people

Just 2–3 years ago, if you told any cybersecurity professional that a company could lose real money because of a deepfake of a CFO, nobody would really believe you or maybe even think that you were stupid. Well, nobody would think that you are stupid now, because there was the Arup case. You probably heard about it; every blog of every security awareness vendor milked that story dry. And it’s easy to think in retrospect that it was predictable, but even if you look now at deepfake technology, using generative AI to create media is still a bit wonky, but when it works, it’s really easy to mistake a deepfake for the real thing. Nowadays, voice deepfakes are a part of a standard toolkit of attackers. And what’s actually interesting is that the realism is just one of the factors and there’s a bit of leeway, because a Harvard-led study found out that it’s not so much how real it sounds, but rather how persuasive the script was. Basically, if the context makes sense, if there’s just enough pressure, people tend to comply, and their internal voice that says that something is wrong is ignored.

There’s also, I guess, an underlying issue of people getting used to email scams, but voice-based phishing in the corporate world was something that didn’t trigger defenses as much. Attackers don’t really need to create voice deepfakes; it’s just a really convenient tool: you don’t have to call yourself and can just deploy an AI agent that can do the talking for you. A good pretext does the heavy lifting. The channel and the manner of attack didn’t change drastically; synthetic media just became another instrument, and the goal is still the same: convince an employee that a weird request is not weird at all.

It’s also worth mentioning that detection basically doesn’t work. First of all, humans are really bad at identifying synthetic media; they are even worse at it when they are under pressure. Hany Farid, who’s spent two decades as a leading digital forensics expert, said in one of the interviews (I forgot which one) that he no longer trusts his eyes. Technical tools that boast 90% detection accuracy in a lab environment drop in accuracy quite rapidly when put in the real world, to where it’s basically a coin toss. But even if we get a tool that can achieve 90% accuracy, it’s actually not good, because, well, what the hell are you going to do with the remaining 10%!? Watermark deepfakes because bad actors are notorious for using commercial tools and following regulations? Give me a break…

Anyway, the point here is really that deepfakes are a real problem. The way cybersecurity professionals deal with it is another question. It’s also worth mentioning that it is not just a technical exploit; it’s basically good old social engineering, augmented by technology.

Housing prices grow, but phishing is cheap

AI made it cheap and good. Not in the sense that phishing is a good thing, but quality is better now. There’s also more phishing now. An experiment by Fred Heiding’s Harvard Kennedy School team found that AI-made emails (this also includes AI-powered research, not just writing) hit a CTR on par with that of emails written by human experts (not Nigerian princes, mind you), at a cost of roughly four cents an email. So, high-quality spear phishing that was reserved for bigger ‘phish’ now can be used for everybody. Economically, it now makes sense.

And even if we somehow develop tools to detect AI writing, it’s still not going to solve the problem. Let’s take writing in general. Now we have tools like Pangram. On the surface, it seems that AI-writing detection is solved, but there are a couple of caveats. On smaller pieces of text, even using an Oxford comma can trigger AI detection. On bigger pieces of text with a bit of manual editing, detectors can be fooled. And even if writing can be detected with 100% accuracy, then what? Are we going to just outright ban the use of AI? And even if some company does it, what if it had to work with another company that doesn’t do it? If we continue this thought experiment, it becomes obvious that an approach like this will fail spectacularly.

Also, using AI for research still gives an enormous edge. AI-writing detectors are trained to recognize AI writing; they cannot stop attackers using AI for research. An attacker can just manually write a phishing email based on the research they did, just as a copywriter can use AI to research the topic, use AI to plan out an article, create a draft, and then just rewrite it from scratch so that there are no traces of AI writing. It would still be faster than doing everything without AI.

Coding is solved, bugs are not

The heading is a quote by Boris Cherny. Why did I use it? Well, if you’ve lived under a rock for the past couple of years, AI is really good at coding. Maybe not as good as humans yet, but good enough. It also creates code faster than a human can write it by hand. Quality might vary, but it is fast. You can also launch multiple parallel agents that will write the code simultaneously. And because AI is good at coding, it’s also good at finding vulnerabilities.

I hope it’s clear where it’s going. Bad actors can use AI to quickly find vulnerabilities and produce malware at scale. Maybe a complete noob will fail, but an attacker with some technical background can use it as a force multiplier. So, attackers can move faster and at a bigger scale than before, so it’s really important to quickly identify issues and patch them quickly, because that tiny gap between a vulnerability being identified and patched is enough. Verizon’s 2026 DBIR, looking at actual malware, found that less than 2.5 percent of AI-assisted samples used advanced techniques. Most of it was trivial, just produced faster.

Again, AI serves as an accelerator. Recon, targeting, writing: all of it just gets faster and cheaper.

Mythos panic

When Mythos came out, it was a huge deal. People were losing their minds and screaming about the end of cybersecurity as we know it. It was the most powerful model ever, and if it fell into the wrong hands, we were done. And to be fair, if we remove all this noise, a model that’s really good at finding vulnerabilities is obviously going to become a dual-use tool. And Anthropic did say that it turned up a whole bunch of high-severity bugs.

The issue here is how the whole thing played out. The model wasn’t released to the public and was released to a limited number of partners: big tech companies, security firms, government-adjacent groups, etc. Then Fable was released and got tangled in an export-control thing.

And here I get a bit cynical. It’s a cool marketing tactic, but what it led to is that access was so restricted that only the select few got to use it. Well, what about smaller companies that still hold user data and have vulnerabilities? Just because a company is not a billion-dollar company doesn’t mean its user data is worthless. And even available models got progressively more restricted; hell, even questions about setting up VPNs sometimes can be blocked. There was an audit that mentioned that refusals of benign cyber requests do exist.

Personally, I don’t have an answer as to what the better approach would be, but I know that criminals are criminals precisely because they tend to ignore the law and regulations. Yeah, they may not get access to Mythos, but they can get access to other models. And they don’t always need to target bigger targets; a smaller company that also doesn’t have access to the best models is on a level playing field with the attacker. So we end up in a situation where access for defenders is rationed, while attackers enjoy democratized access through open-weights models.

But there’s yet another catch! The human element is responsible for about 62% of breaches based on Verizon’s data. What does it mean? Well, patching all vulnerabilities still doesn’t make you 100% secure! Convincing an employee to give out an access token doesn’t require finding vulnerabilities. Again, I don’t pretend to know that I know better, but I do know that refusal even from simple models to help with self-hosted VPN setup probably is not the way to go forward.

So, AI did have an impact here, but again the reality is slightly more nuanced and complex.

AI for security awareness training

Let’s look more closely at how AI is used by defenders.

The security awareness training industry has a wonky record. There was research of all kinds showing that courses don’t really work, or at best the impact is very small. Let’s be honest, passively watching a 7-minute video about GDPR is nobody’s idea of a good time, especially if you have to do it when you have a bunch of meetings, sales calls, and reports to finish. It’s just a compliance checkbox. Honestly, I don’t know why research was needed to figure that one out.

So, what’s the answer? Generating more content with AI. And I actually think it’s a good idea if done in a very specific way, because there are benefits to it. First of all, it’s scale. You can generate a bunch of courses very fast that cover all kinds of topics. You can also generate courses for very specific cases or tailor them to specific employees. Or if a new type of attack appears, you can quickly create a course that trains employees on this specific attack. But it doesn’t mean it’s going to be good.

I actually wrote my fair share of cybersecurity courses, and it’s hard. To create a good and engaging course is not easy. You have to know the topic very well; you need to understand how to progressively disclose the information so that the reader will not get lost; and you have to figure out how to engage the learner and keep their attention. It’s really hard. What’s even harder is to explain to any AI model how to do that, because (I’m not sure if you know this) AI cannot think and reason as a human does; it doesn’t have taste. AI doesn’t really know where to put a joke in; it doesn’t understand if cognitive load gets too high; it doesn’t really know how to effectively explain complex topics. So, training AI to do that is not an easy task.

I saw some SAT vendors trying to do that. Adaptive Security had a free course on deepfakes, which was clearly AI-generated. There were some very good parts in there, but there were also very questionable parts there that seemed like very low effort. That’s where human authorship still has the edge, but things are moving forward. AI can be a powerful education tool; it just needs a very clear description of what good is, what pedagogical rules to follow, and other guardrails that prevent it from drifting into slop. And it means that you have to give it a lot more structure than you would ever give to a competent human. This is one of the areas where scale doesn’t really help, and vendors should focus on quality.

Where AI actually helps defenders

I think one of the obvious wins is simulations. First of all, there’s vishing. There are vendors now that provide vishing simulations that use AI agents, because they are now good enough not only to convincingly copy voices of real people but also to hold a conversation and adapt in real time. Now you can run vishing simulations at scale and train employees to recognize voice phishing and train them to use proper verification methods and stuff like that. That’s actually really cool.

Email simulations also benefit from AI. AI can be used to personalize simulations, where each employee can get an email that is actually tailored to their role and that makes sense to them. AI can choose the right templates, fill in the right data, or write emails from scratch. Simulations can be updated quickly to mirror real attacks.

There’s a real positive gain here, where rehearsal can become very effective if done right and can help employees recognize social engineering. I mean, it’s a bit clearer how AI can be used on the technical side of things; with social engineering, it’s a bit less clear, but it seems that the industry actually found a good use of AI in this area.

I think I’m not going to spend a lot of time outlining the benefits of AI when it comes to vulnerabilities; it’s basically self-evident from previous sections of this article. Just as attackers can find vulnerabilities and write malware, defenders can find the same vulnerabilities and produce patches faster. AI can help analysts; it can help prioritize which vulnerabilities to patch first, map which assets are actually exposed, and do all that jazz. It’s still not perfect and requires human oversight, but as with criminals, AI can become a force multiplier for defenders.

So, did AI cause a cyber-apocalypse?

Ehh, not really. I mean, just yesterday, August 27, 2026, there was news about some hackers using Cursor to hack multiple companies. Details are not entirely clear, because it seems that most reporting focuses on the fact that criminals tricked one of the models. Details of exactly what happened and how are not yet clear at the time of writing this article.

Then there’s this open call by OpenAI for collective action on cyber defense… I mean, yes, things are going to escalate. Criminals will try to find better ways of using AI, but defenders will too. That’s a very expected and logical assumption; I’m not sure why another call for action is needed, but I hope it will result in something useful.

Anyway, AI did change cybersecurity; it is also changing security awareness, and it will continue to change things, but I don’t think we actually will see extreme scenarios. The reality of our world is that extreme predictions of extreme events tend to be overexaggerated, because humans are actually bad at predicting things; even with enough data, it’s still really hard to predict what’s going to happen, but monitoring the situation in this field is a really good idea. Also, be ready to move fast; one clear, noticeable trend is that AI accelerates things, so try to keep up.