Why the next hacker might not be human

In the first episode, we discuss how artificial intelligence is changing vulnerability discovery and why it has become both the primary defender and a new tool for attackers.

Abbos Rikhsiev, Tashkent, Chief Growth Officer at Cerberus, LinkedIn

How AI learned to find holes in other people's software and does it better than we do.

Just a few years ago, finding vulnerabilities in software was considered work for a narrow circle of specialists. Only those who had spent years reading other people's code and had a deep understanding of system architecture could find a working hole in popular software. In 2025 and 2026, this barrier collapsed. AI learned to find vulnerabilities independently, without human supervision, in a matter of minutes and at almost no cost, and today an increasing number of companies are having to rethink their approaches to cybersecurity.

The long history of one idea

The dream of a program that could find holes in other people's code on its own did not appear yesterday. As early as April 1995, the SATAN scanner — Security Administrator’s Tool for Analyzing Networks — was released on the internet. Its appearance caused a stir at the time: newspapers and television wrote about it as a nearly ominous invention capable of handing over any network to outsiders.

In 1998, Nessus appeared, followed by the Russian XSpider. These programs could scan networks and check them for known issues: open ports, outdated software versions, and typical configuration errors. Essentially, these were huge lists of ready-made rules that compared the system against a database of vulnerabilities already described by others.

But none of these tools could do the main thing — find a fundamentally new vulnerability that no one had known about before. For this, a program would need to "understand" code logic roughly as well as an experienced developer does. Just a few years ago, such an idea was not taken seriously — it sounded like science fiction.

A turning point emerged when language models had just learned to confidently write emails, create presentations, and conduct correspondence. The thought that similar technology could independently read other people's source code and find logical errors in it still seemed like an exaggeration. Most security specialists were skeptical about it.

The situation began to change with the arrival of next-generation models, such as GPT-5 and Claude, capable not just of answering questions, but of reasoning step by step and keeping thousands of lines of code in their field of attention simultaneously. Early experiments quickly showed that the task was not only solvable, but solved significantly faster than the developers themselves had expected.

Where it all began

The first major incident occurred in October 2024. Google launched the Big Sleep experiment to test whether AI could independently find a vulnerability in a real program. For the test, they chose SQLite — one of the most common databases in the world.

Just a few days later, the model discovered a previously unknown vulnerability. A patch was released on the same day, and the media called it "the first case in history."

A year and a half passed — and what then seemed like a sensation became ordinary practice. Today, AI regularly finds new vulnerabilities faster than humans, and there are already hundreds of such cases.

The Firefox browser story

Anthropic, the developer of the Claude language model, partnered with Mozilla, the creators of the Firefox browser, to conduct a joint experiment. The model was given access to the browser's source code and began studying it on its own, without human supervision.

It found its first serious vulnerability in 20 minutes. While people from Anthropic were writing the report and verifying the finding, the AI discovered another half-dozen suspicious areas. Over two weeks, the model reviewed about 6,000 files. Mozilla ultimately confirmed 22 real vulnerabilities, 14 of which were critical. This is almost one-fifth of all critical holes that Firefox fixed during the entire previous year.

One of the vulnerabilities was rated by experts as maximally dangerous — nearly 10 out of 10 on the international severity scale. Through it, a hypothetical attacker could take over someone else's computer. It was also found not by a human.

Especially impressive is the price. The entire process — searching, verifying, and writing reports — cost approximately $4,000. Less than it would have cost to hire one bug hunter for two weeks.

The competition in Las Vegas

Six months before the Firefox story, in the summer of 2025, the American defense agency DARPA organized a competition that sounded like a science fiction plot. Seven teams developed autonomous digital vulnerability hunters. The task was to find as many holes as possible in critical software used in the real world: in the Linux operating system kernel, in databases, and in utilities responsible for transmitting data over the internet.

The total volume of code the systems scanned was 54 million lines. That's roughly the same amount as all of Russian classic literature if you add up Tolstoy, Dostoevsky, Chekhov, and Gogol, and then multiply by twenty. In 143 hours of autonomous work, without a single command from a human, participants found eighteen previously unknown vulnerabilities in real software.

The winner received $4,000. The average cost per task processed was $152. DARPA Director Steven Winchell commented on this without pomp: "We live in a world where digital infrastructure rests on ancient wooden scaffolding. A problem of this scale can no longer be solved by a human alone."

Who is already using it

Big Sleep from the experiment has long become a working tool for Google. In the summer of 2025, it found a hole in the same SQLite before attackers could exploit it. According to the company, the attack was already being prepared — but it was preempted. For the first time in history, AI did not react to a cyberattack, but prevented it.

OpenAI presented its own digital cybersecurity specialist with the charming name Aardvark in the fall of that same year. In just a few months of operation, he received official recognition for discovering a dozen serious vulnerabilities in open-source software. The startup XBOW, with its AI, took first place in the American HackerOne bug bounty rankings, surpassing all living competitors.

A separate story involves the cURL library. It is 30 years old and sits inside almost every internet-connected device. Its lead developer, Daniel Stenberg, told American radio station NPR in April 2026 a simple fact: in the first three months of the year, his team fixed more vulnerabilities than in each of the two previous years combined. Some of them — and he himself actively uses AI — he found precisely this way.

For example, on the Cerberus platform, the world's first secure AI hacker, a proprietary language called Cerberus Lang is used, through which system actions are described and permissible testing boundaries are set. According to developers, this approach allows recording every step of the agent and limiting the execution of operations that go beyond the agreed scenario. Such mechanisms become especially important as autonomous tools begin to gain access not only to source code but also to live corporate systems.

Another illustrative story happened with OpenSSL. This is a library that ensures encryption for a significant portion of internet traffic — from online banking to government websites. In January 2026, a scheduled release came out with fixes for 12 vulnerabilities. An autonomous system called AISLE, unrelated to the OpenSSL team, found all 12 in the same timeframe — plus several historical bugs that had quietly lived in the code for years. Recall, this is one of the most thoroughly checked programs in the world, studied for years by thousands of security specialists.

The flip side

Up to this point, the plot sounds optimistic. White-hat researchers, responsible disclosure, everything under control. But the models on both sides — for defenders and attackers — are the same. And here the picture changes sharply.

In the fall of 2025, the same Anthropic recorded the first documented cyberespionage campaign in which AI took on the role of the operator. Humans set the goal, and the model carried out the multi-stage attack almost independently. According to IBM, one in six data breaches registered in 2025 is already linked to the use of AI by the attacking side. Sending convincing phishing emails, fake videos and voices, automatic password guessing.

The picture in Russia is similar. According to Kaspersky Lab, one in three cyberattacks on companies in 2026 begins with the exploitation of software vulnerabilities. The total number of holes in programs grew by almost a quarter over the year. The company itself admitted that it now runs all its source code through its own AI model — despite having hundreds of live security experts working there.

For decades, the industry relied on an unwritten assumption: to find a working vulnerability in a serious program, you need years of experience, specialized knowledge, and a lot of time. The circle of those capable of doing this was narrow precisely because the entry barrier was high. Well, that barrier has just collapsed. Its height is now determined not by human expertise, but by the cost of machine time.

What should people do

While model developers and corporate security services run a race, ordinary people have one extremely simple rule. Install updates as soon as they arrive. All of them, everywhere, on smartphones, laptops, routers, and set-top boxes. Previously, days or even weeks passed between the release of a patch and the first attempt by hackers to exploit it. Users had time. Now, there is no such time.

For companies, the rule is slightly more complex, but fundamentally the same. The difference between "we do manual code review" and "we run every significant change through a language model" will soon turn into the difference between protection and its absence. Not because a live specialist is worse than a machine. But because at such a pace, no live specialist will be able to read everything that needs to be read.

A human hacker, of course, is not going anywhere. A specialist who knows a specific target, knows how to exploit human weaknesses, and can make non-standard decisions will still be part of real scenarios for a long time. But everything that used to make humans unique in this profession — patience, hours of reading other people's code, searching for non-obvious moves — is now being done by models. Sometimes better. Almost always faster.

The real question now is not whether the next major hack will be done by AI. It already is. The question is who will get there first — the attacker or the defender. And every company is currently answering this question for itself. Even those who have not yet realized it.