The Machines Are Now Searching for Vulnerabilities Themselves - And the First Real Warning Is Already Here

For years it sounded like one of those futuristic warnings from cybersecurity conferences that people briefly read and then forgot. Someday criminal hackers would use artificial intelligence to independently discover security flaws no human had ever found before. Someday programs would begin searching for the internet’s weaknesses on their own. That moment now appears to have arrived. Google published a report on Monday that is causing significant alarm throughout the cybersecurity industry. According to the company, a criminal hacking group attempted to prepare a large-scale cyberattack with the help of an AI model. The target was apparently a previously unknown vulnerability in a widely used open-source management system. Such weaknesses are considered especially dangerous in the industry because even the developers themselves are unaware of them. That is why they are known as so-called zero-day vulnerabilities.

According to Google, the attackers used a model not only to identify the vulnerability but also to directly prepare it for an attack. Security researchers have warned about exactly this scenario for years. Until now, the search for such vulnerabilities was considered extremely labor-intensive. Specialized groups often spent months working on individual systems. Now part of that work may be becoming automated. What is especially remarkable is not even the technical side itself, but the speed at which the situation is changing. Only a few years ago, zero-day vulnerabilities were considered rare commodities capable of fetching millions of dollars on illegal markets. Now cybersecurity firms openly say that modern AI systems can uncover vulnerabilities on a massive scale.

The pressure intensified further after the company Anthropic unveiled its new model Mythos. According to Anthropic, the system identified thousands of unknown vulnerabilities in major operating systems and internet browsers - including flaws that may have remained undiscovered for decades. Precisely for that reason, the model was not publicly released, but instead made available only to selected companies and government agencies in the United States and the United Kingdom. Google says it detected the recent attack in time. The vulnerability was patched before major damage could occur. However, the company has not disclosed the affected software. The responsible hacking group also officially remains unknown. The only confirmed detail is that the attack apparently used a Python script and aimed to bypass two-factor authentication. To do so, the attackers would additionally have needed valid credentials such as usernames or passwords.

John Hultquist of the Google Threat Intelligence Group nevertheless describes the incident as a turning point. He said it was the first concrete case in which there is high confidence that artificial intelligence was actively used to discover and prepare a previously unknown vulnerability. In his words, this is likely only “the tip of the iceberg.” Particularly disturbing is how Google claims to have identified the use of AI. According to the report, the code contained unusual explanatory comments and phrasing that experienced human developers would normally never include. Rob Joyce, former cybersecurity director at the NSA, explained that AI-generated code now leaves traces that can sometimes be identified - even if the machines do not openly mark their origin.

While governments are still arguing over how AI should be regulated, reality is already changing faster than the political debate. Security agencies and technology companies are now openly discussing whether new models should only be released under controlled conditions. The reason is fear that highly advanced systems will not only help researchers, but simultaneously provide cybercriminals with tools that previously existed only within intelligence agencies or specialized groups. But the real problem runs deeper. Today’s internet was built over decades by humans - with mistakes, compromises and security flaws. Those accumulated weaknesses are now colliding with systems capable of analyzing millions of lines of code in seconds. The machines are no longer just learning how to write. They are now learning where existing systems are vulnerable.

For the moment, much still appears controllable. Google managed to stop the attack. The vulnerability was patched. But that is precisely where the real warning lies. If this truly is only the beginning, then the balance of power between defense and attack in the digital world is changing right now. Not sometime ten years from now. Right now.