Here is the stat that should keep every CISO awake tonight: in the first confirmed AI-driven cyberattack, human hackers handled only 10 to 20 percent of the work. The AI agent did the rest.

Anthropic disclosed late last year that state-sponsored Chinese hackers used its AI technology to infiltrate roughly 30 companies and government agencies worldwide. In a blog post, the company called it the first reported case of a cyberattack where AI gathered sensitive information with limited human involvement. Five months later, it remains the only confirmed example. But cybersecurity experts are now warning, loudly, that the window of relative safety is closing fast.

The New York Times reported this week that as Anthropic, OpenAI, and Google prepare to release their next generation of AI systems, the cybersecurity community is sounding alarms. These models could allow hackers to identify security vulnerabilities in computer systems far faster than any human team. The decades-long arms race between attackers and defenders is about to get a turbo boost, and the attackers have the advantage of moving first.

Think about what this means in practice. Right now, finding a zero-day exploit in a major piece of software takes skilled hackers weeks or months of manual probing. An AI agent with the right training could do it in hours. It could scan millions of lines of code, identify patterns that match known vulnerability classes, and generate working exploit code before the coffee gets cold.

The silver lining, if you can call it that, is that defense gets the same upgrade. AI-powered security tools can monitor networks in real time, detect anomalous behavior patterns, and patch vulnerabilities faster than human teams ever could. The question is whether defenders will adopt these tools as fast as attackers will.

History says no. Attackers are always faster because they only need to find one hole. Defenders have to cover every surface.

The uncomfortable truth nobody in the AI industry wants to say out loud: every capability improvement in frontier models is also a capability improvement for attackers. When Anthropic and OpenAI brag about their models' ability to write code, reason through complex problems, and use software tools autonomously, they are describing exactly the skills a cyberattacker needs. The same AI that writes your marketing copy can write your malware.

We are entering the era where AI does not just help hackers. AI IS the hacker. And the 30 companies hit in that first attack were just the proof of concept.