AI broke the cybersecurity world. What’s next?

In a Decoder interview with The Verge, Alex Stamos, chief security officer at Cognition and a lecturer at Stanford University's Cyber Policy Center, described October as a continuous emergency for defenders and argued the cybersecurity industry built up since the early 2000s is being outpaced by AI-enabled attacks. He called the past weeks "security awareness month" in a grim sense: alongside widely reported bugs in PHP and Node, he said a large number of companies were breached, including a succession of AI-focused neoclouds.
The concrete example he gave shows how these incidents compound. According to Stamos, attackers broke into Fireworks, a prominent provider of AI services, and stole API keys that let them get into Cursor. From Cursor they pushed a bad plugin that reached 5,000 more companies, which he described as 5,000 potential breaches. He also cited the Binance hack, which he characterized as the largest heist of anything in human history, and said a CISO friend asking colleagues "Is anybody else tired?" drew unanimous agreement because teams had effectively been at DEFCON 2 through October and into November. (DEFCON levels are the US military's readiness scale; DEFCON 2 is one step short of war.)
On attribution, Stamos said the actors are varied rather than a single group. He attributed the Binance theft to North Koreans, describing a dedicated team whose work is stealing Bitcoin, and said that operation leaned mostly on human intelligence and social engineering rather than AI. The breaches of AI companies he tied to what he called the Chinese cottage access industry: small Chinese contractors who break in and sell what they find to much larger Chinese companies.
The mechanism he considers most important is an inversion of the cyber "kill chain," a term borrowed from the military for the steps taken to have an effect on a target, such as stealing information or money. Traditionally, attackers did research and exploit development well ahead of time, then deployed that finished work against one target or against hundreds or thousands in rapid succession. With AI, Stamos said, a system can examine targets, find bugs and write exploits in real time instead. The tool drawing attention is ARTEX, a Chinese open-source project with an English translation on GitHub that anyone can inspect and which he said, run on a backend model like an uncensored GLM-5.3, can break into a striking number of organizations. He said the campaign began with three South Korean banks and that roughly 200 more companies were hit in the week of the interview.
Stamos framed the current period as a return to the 1990s, when early DEFCON gatherings drew a few hundred people in a single track and a talk could convince attendees that every company in the world was exposed. The difference now, he said, is that the world runs on the internet and attacks can be automated. He also pointed to the disappearance of the old skill hierarchy, in which a small tier of elite vulnerability researchers sat above script kiddies who merely ran other people's tools. He said he can now reverse-engineer a Microsoft patch and write stable exploit code on a machine sitting on his desk; it is not yet fully automatic and still requires real skill, but it is close. The operational side, he added, is where AI lets an attacker scale.
His account of the previous security era starts with Bill Gates's "trustworthy computing" memo in the early 2000s, when large vendors began hiring former hackers and the industry developed responsible disclosure, bug bounties, transparency reports and consultancies that grew into firms such as Mandiant and CrowdStrike. That apparatus rested on several assumptions now eroding: a limited supply of skilled people, a conspiracy of specialists needed to run an attack team, and the need for money launderers, the last of which Bitcoin undercut. Stamos argued defenders were never actually secure and that the past decade of ransomware mostly hit midsize firms few have heard of but many depend on, down to a medical and dental supply distributor whose disruption left dental offices canceling appointments over missing gloves.
His prescription is economic. Attacks are becoming agentic, he said, and the per-token cost for attackers is trending toward zero while defensive costs stay high, so the priority is to make defense cheaper, something he said few people are focused on. He estimated software security is only about 25 percent of a CISO's job, and that AI has revolutionized that quarter while helping the rest only somewhat and in some cases making it worse. The next frontier he named is security operations: ingesting terabytes or petabytes of logs with cheap traditional classifiers such as TF-IDF models and Bert variants, escalating anomalies to a small language model, which can call a larger one, and eventually replacing first- and second-tier human analysts with AI. That layered design is a practical reading of how to keep inference costs bearable.
Why it matters: Teams shipping software and AI services now inherit risk from their suppliers, since one stolen API key can propagate through a plugin to thousands of downstream customers, so product and platform teams should treat third-party integrations and key scoping as part of their own perimeter. The shift toward cheap open-weight models on commodity hardware suggests AI-assisted intrusion tooling will spread faster than the defensive tooling built to match it. Security budgets and hiring may need to move from bug-finding toward detection and response automation, and buyers should expect vendors to explain how their pipelines handle compromised dependencies. That last point is inference from Stamos's account, not a claim he made about specific vendors.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Today, I’m talking with Alex Stamos, chief security officer at Cognition and longtime cybersecurity expert, about what a mess cybersecurity is right now. This episode is part one of a special series looking at AI in real life. I’ve always enjoyed talking to Alex because he’s so direct, and you’re going to hear him say […]