Alex Stamos says AI has inverted the cyber kill chain

The Verge published an interview with Alex Stamos on its Decoder podcast on Nov. 7, 2026. Stamos is chief security officer at Cognition and a lecturer at Stanford University Cyber Policy Center. He said the multibillion-dollar cybersecurity industry is overdue for waves of AI-boosted hacks, breaches, and attacks, and that the industry is entering a dark age while it figures out how to adapt to AI.

Stamos compared the current environment to the 1990s frontier feel when the always-online world was brand-new. He traced the prior security era back to Microsoft’s Bill Gates trustworthy computing memo in the early 2000s. That era saw Microsoft, Oracle, Amazon, and Salesforce hire ex-hackers, which led to responsible disclosure, bug bounties, and firms such as Foundstone, @stake, Mandiant, and CrowdStrike. Stamos noted that early DEFCONs in the 1990s had a couple of hundred people and one track, unlike today’s 40,000 to 60,000 person event.

This model assumed a skill hierarchy from script kiddies to a god tier of hundreds of elite vulnerability researchers, and that only a limited number of people could be operational in an attack, with elite R&D and larger operational teams. Professional teams at U.S. Cyber Command, Russia SVR, and Chinese MSS used that structure, a model later copied by Russian ransomware cartels. Stamos said Bitcoin changed the economics because it let nerds launder money. In the 1990s, the entire world did not run on the internet, so attackers could not automate ruin at scale.

Stamos said open-weight AI models can now automate most of the cyber kill chain. The kill chain is the sequence of steps an attacker takes to breach a system. Traditionally, attackers researched software, found vulnerabilities, wrote exploits ahead of time, and humans deployed them against one or thousands of targets. With AI, the kill chain has inverted, so systems find bugs in real time, write exploits in real time, and break into companies.

He demonstrated the accessibility of this capability by running a quantized, obliterated GLM-5.3 model on an M5 Mac Ultra Studio, a version compressed for consumer hardware and stripped of its safety controls. Attackers no longer need frontier models with protections and logs available to law enforcement. AI lets script kiddies use much of the skill set of elite researchers. On a desktop Mac, someone can reverse engineer a Microsoft patch and write stable exploit code, though it is not fully automatic yet.

Stamos said two or three Russian kids can now run a bank of Macs or buy Nvidia H-100s, and they can run agents all day rather than needing a whole team. This shift is evident in recent attacks: over the past month, there were major bugs in PHP and Node, and a huge number of companies were breached. Neoclouds, which are AI-related companies, have been breached in succession, and Stamos called the Binance hack the largest theft ever.

The Binance attack illustrates the human element. Stamos said that for Binance, North Koreans run a team focused on stealing Bitcoin, using mostly human intelligence and social engineering. In another incident, attackers stole API keys from Fireworks, an AI services provider, and used them to break into Cursor. From Cursor, the attackers pushed a bad plugin to 5,000 more companies, causing 5,000 potential breaches.

Another tool showing the automation trend is ARTEX, a Chinese open-source tool with an English translation on GitHub. Stamos said ARTEX can break into many organizations when powered by an obliterated GLM-5.3 model. ARTEX attacks started with three South Korean banks, and about 200 more companies were hit this week. Stamos said that downloading an open-source tool to break into multi-billion dollar companies has not been possible for a long time.

Stamos noted that October is traditionally security awareness month, but a CISO friend of his said the industry has effectively been at DEFCON 2 during October and into November. Stamos noted that attacks are being carried out by a variety of groups, not just because adversaries decided to act. He summed up the situation by saying, “It’s a variety of different groups.”

He said private actors and state actors including Russia and North Korea are breaking into systems with AI helping them detect and exploit vulnerabilities simultaneously. Attackers now include a larger ecosystem of people in places law enforcement cannot touch, with economic incentives and access to tools. Russian ransomware cartels have hit mid-size companies, including the country’s largest distributor of medical and dental supplies, disrupting dental offices, and that medical supplier had one security guy and a 10 to 12 person IT team facing a 20-person Russian ransomware cartel.

Breakers of AI companies are part of the Chinese cottage access industry, where small contractors breach companies and sell information to larger Chinese companies. Stamos said attacks are becoming agentic, meaning they can operate autonomously, and attacker token costs are falling toward zero while defender costs remain high. Defenders need cheaper options, and he is personally focused on lowering defense costs.

At Cognition, where he is chief security officer, software security is 25% of his job, and it is the part most revolutionized and improved by AI. The other 75% of his job has been helped a bit or made worse. The next focus area should be operational security, such as watching logs and detecting techniques against the company. Companies must respond without shutting down the business.

Stamos argued that companies that want to survive must replace first-tier and eventually second-tier human SOC workers, the teams that monitor and respond to security alerts, with AI using current LLMs. However, the first tier of log analysis cannot be an LLM because scanning terabytes or petabytes with an LLM is uneconomical. Simpler tools such as TF-IDFs, which identify important words in logs, and MLBERTs, which classify security events, should flag issues first.

A colleague at a large security platform company said one investigation on a frontier cyber model costs $900 on their product. Russian ransomware cartels using Macs or throwaway crypto mining rigs have near-zero marginal cost, making $900 investigations unworkable. Cognition trains its own models, retrains open weights, and is trying to drive investigation costs down.

The industry must focus on the Pareto frontier, the set of the most efficient trade-offs between cost and efficiency in cybersecurity, not just bug-finding. Mythos and Astra Cyber are strong models but only a tiny number of companies can access them, and they scan only the most important code. Code outside the Linux kernel runs to hundreds of billions of lines, including older C# and J2E code from firms like Infosys, and expensive frontier cyber models will not scan that code. Stamos argued that lowering the cost of scanning that legacy code is essential, and he said he is neither the hardest doomer nor the most positive, and does not believe everyone is doomed.