AI broke the cybersecurity world. Now what?

Alex Stamos says October’s usual Security Awareness Month message—change passwords and use a password manager—was overwhelmed by a succession of serious incidents. AI companies and “neoclouds” were breached, Binance suffered what he calls the largest heist in human history, and the Cursor incident showed how attacks can chain together: stolen Fireworks API keys enabled access to Cursor, whose malicious plugin reached 5,000 companies and created thousands of potential breaches. A CISO friend asked whether everyone was tired; after staying at “DEFCON 2” through October and into November, everyone answered yes.

Stamos says open-weight models have reached the point where they can automate most of the cyber “kill chain”—the steps used to steal information, shut systems down, or take money. Instead of humans researching a target, finding a vulnerability, writing an exploit, and then deploying it, an AI system can inspect targets, discover bugs, write exploits, and attack in real time. He points to Artex, a Chinese open-source tool with an English GitHub translation, which, paired with an uncensored model such as GLM 5.3, reportedly broke into three South Korean banks and then roughly 200 more companies.

Stamos compares the moment to the late 1990s, when a few hackers could download tools and discover that nearly every company was vulnerable—except the internet was not yet central to everything. The industry later built responsible disclosure, bug bounties, security consultancies, and large defensive companies around a hierarchy separating script kiddies, tool writers, vulnerability researchers, and elite “gods.” Bitcoin already changed the economics by letting technically skilled criminals launder money; AI now gives script kiddies capabilities once limited to elite researchers and lets a few attackers run many automated operations against companies.

The host argues that AI gives attackers real incentives and capabilities, not merely more teenage mischief. Stamos agrees: attacks are becoming agentic, attackers’ cost per token is approaching zero, and defensive costs remain high, so the priority must be lowering the cost of defense. Bug finding is only about 25% of his CISO job; the other 75% includes security operations, log analysis, investigations, and response. He proposes cheap classifiers for enormous log volumes, followed by small and larger language models, because one investigation using a frontier cyber model reportedly costs a security company $900.

Stamos rejects both immediate extinction and the idea that security will become permanently perfect. He expects a “dark age” in which breaches continue even when software has no flaw, and flaws remain even if language models write code in memory-safe languages. The eventual median codebase may improve, but decades of unsafe human-written code remain. When a codebase contains 10,000 bugs, he says rewriting it with extensive tests and specifications can be cheaper than patching every flaw. He cites recurring new vulnerability pools found by Mythos and AstroCyber in the Linux kernel as evidence that scanning alone is not yet winning.

Stamos urges Google and Amazon to spend perhaps $100 million to $1 billion rewriting the Linux kernel in Rust or another type-safe, memory-safe language, focusing first on production server hardware rather than obsolete drivers. He says a fork could remain ABI-compatible with Glibc and eventually replace parts of that stack, allowing cloud providers to migrate without waiting for Linus Torvalds to abandon support for old hardware. LLMs could also write contracts describing how functions should behave and generate tests. The goal is to stop repeatedly patching infrastructure whose exploitation timelines are now faster than companies’ patch processes.

Stamos says every company should “R-TEX” itself within 90 days: run an AI attack or an equivalent red-team product against its own systems. He expects a wave of breaches by attackers running Artex locally on a gaming computer or Mac, especially against forgotten websites and API endpoints. He acknowledges that AI red-team products are nondeterministic and their contracts may disclaim catastrophic consequences, but says companies must accept that risk because an authorized test is less damaging than criminals stealing data, demanding ransom, or shutting down critical systems. Existing legal and vendor-risk processes are too slow to permit this.

Stamos dismisses “swarm” as an overused term: the Hugging Face attack involved roughly 700 to 1,400 agents and perhaps $300 million to $1 billion in hardware, whereas Artex attacks likely use one or two agents. He supports AI handling early defensive actions because organizations generate terabytes or petabytes of logs, but wants deterministic limits and human approval for consequential steps. Cognition’s SOC, built on Devin, may disable Okta accounts but only up to a preset number; exceeding that limit pages a human. He says autonomous bug fixing is safer to deploy first, with models scanning repositories, writing tests, and automatically merging verified fixes.

On Instagram’s AI-generated media, Stamos says Meta can overcome technical objections when it truly wants to, and argues that detecting fakes will remain probabilistic while verifying authentic media can be cryptographically reliable. He suggests capturing media directly from a phone camera, restricting edits to in-app tools, and maintaining the chain of custody inside the phone and platform. His teenagers dislike AI slop, so he predicts a new social app could disrupt Instagram by guaranteeing camera-originated images. He also warns that frontier labs’ safety improvements mainly address models escaping or attacking directly; the larger threat is criminals using models from many sources.

Asked for hopeful indicators, Stamos predicts AI will enable many tiny companies and help individuals build products that once required 50 or 100 employees and substantial venture capital. For cybersecurity, he wants to see fewer major breaches, fewer wormable vulnerabilities in foundational components such as the Linux kernel, OpenSSL, and Glibc, and a decline in Patch Tuesday counts after they rose from roughly 150–200 fixes to 900 or more. He proposes a “Dark Age Index” tracking those measures, especially whether each new frontier cyber model produces a smaller wave of anonymous Linux-kernel and infrastructure patches.