Cognition chief security officer Alex Stamos told The Verge on November 7, 2026 that AI has ended the era when cybersecurity economics favored defenders. In a Decoder interview with Nilay Patel, he said open-weight models can automate most of a cyber kill chain on hardware anyone can buy, that breach chains like stolen neocloud keys flowing into IDE plugins are now routine, and that the industry is entering a "dark age" until logging, triage, and legacy code are rebuilt for agent-speed attacks. This post unpacks what he said, what matches public incident reporting, and what practitioners can do this month.
TL;DR: Stamos's core claims
| Topic | What Stamos said on Decoder | Why it matters |
|---|---|---|
| Capability takeoff | Open-weight models let attackers automate kill-chain steps in real time | Shrinks the skilled-attacker bottleneck |
| October stress | CISOs describe operating at DEFCON 2 since Black October | Incident response capacity is saturated |
| Breach chaining | Fireworks keys led to Cursor compromise and a bad plugin at scale | Supply-chain keys became lateral movement |
| Economics | Attacker token cost trends toward zero; one frontier SOC investigation can cost $900 | Defense must get cheaper, not just smarter |
| Dark age | Painful transition, not doom; shorten it by rewriting unsafe stacks | Patching alone will not keep pace |
What Stamos says changed in the last month
Cybersecurity zero-day concept art for AI-accelerated breach chains Stamos describes on Decoder
Patel opens the episode by calling the stretch from October into November a period when "the multibillion-dollar cybersecurity industry as we've known it is essentially overdue" to AI-accelerated attacks. Stamos answers with a lay of the land tied to both disclosure season and live breaches.
He points to major software flaws such as PHP and Node.js, but stresses that a huge number of companies have been breached even when those bugs are not the headline. He names AI infrastructure providers—neoclouds—falling in sequence, the Binance hack as the largest heist in history, as Stamos puts it on Decoder, and a Cursor incident where compromise at one vendor cascaded outward.
The Cursor chain, as Stamos tells it on Decoder, started when attackers broke into Fireworks, stole API keys, used them to enter Cursor, and pushed a malicious plugin that reached on the order of five thousand companies. That pattern matches what security teams have been warning about in our coverage of ARTEX-driven campaigns and agentic breach reporting: credentials at one AI supplier become keys to many downstream customers.
Stamos says a CISO friend asked a room full of peers if anyone was tired and everyone said yes, because the field has felt like DEFCON 2 through October and into November. That is anecdotal, but it lines up with public exhaustion around Check Point's ARTEX operator logs and the wider OpenAI–Hacktron testing partnership narrative about offensive models running ahead of patch velocity.
Kill chains, open weights, and the return of the Wild West
Stamos defines a cyber kill chain as the steps an attacker takes to affect a target—steal data, deploy ransomware, or move funds. Historically, teams built exploits ahead of time and humans deployed them in bursts. AI inverts that, he says: systems can look at live targets, find bugs, write exploits, and break in in real time.
He attributes much of the shift to open-weight models run on commodity hardware. On Decoder he mentions running a quantized model on an M5 Mac Ultra Studio and the leverage that gives an individual operator. Nation states, North Korean Bitcoin teams, and Chinese contractor ecosystems still matter, but the script-kiddie tier gains skills that used to sit with a few hundred elite researchers.
The tool name he repeats is ARTEX, described as a Chinese open-source framework anyone can inspect on GitHub, backed by powerful models in operator setups. That aligns with Check Point Research's November report, which counted 1,608 live ARTEX panels among 2,557 tracked systems and near-autonomous operator campaigns. Stamos compares the moment to late-1990s DEFCON, when a single talk could make every company feel exposed—except today the whole economy runs online and automation scales abuse.
Patel and Stamos also discuss who is behind attacks: North Korea on Binance-style thefts, Chinese cottage-industry contractors selling access, and Russian ransomware teams that used to need large rosters. AI, Stamos argues, promotes every attacker a league, much like European football relegation—more actors can play at a higher tier.
Economics: cheap attacks, expensive investigations
A long section of the interview is deliberately economic. Stamos says the Microsoft Trustworthy Computing era assumed a skill hierarchy and limited operational headcount. Bitcoin removed money-laundering bottlenecks; AI removes research and exploitation bottlenecks.
On defense, he warns that industry attention over-indexes on AI bug finding. As Cognition's CISO, he says software security is only about a quarter of his job; the rest includes security operations center work—reading logs, deciding whether to page people, and responding without shutting the company down.
Stamos describes a tiered SOC architecture: dumb classifiers on terabytes of logs, then small models, then larger ones on escalation. He quotes a colleague at a large security platform who said a single investigation on a frontier cyber model can cost around nine hundred dollars on their product. Attackers running local models on Macs or cheap GPU rigs face near-zero marginal cost. Closing that gap, he says, is something Cognition is working on by retraining open weights, but "very few people are" focused on defender cost curves.
That is the bridge from interview theory to November headlines. When Chainalysis widened its breach disclosure and attributed techniques to offensive agentic methods, it illustrated the same asymmetry: automated offense against human-paced customer response.
The dark age and rewriting the internet's unsafe code
Patel asks whether Stamos still believes cybersecurity is entering a dark age. Stamos says yes, but qualifies the metaphor: dark ages end, and defenders must shorten this one—protect the "Library of Alexandria," as he puts it on Decoder, by not relying on endless reactive patching.
His proactive prescription on the show overlaps with themes he voiced on Galaxy Brain in late October with Charlie Warzel, as summarized by Outside the Beltway's recap of that interview: memory-unsafe C and C++ stacks accumulated decades of flaws; frontier and open models now burn through those bugs faster than patches deploy worldwide. Stamos argues some critical software must be refactored in Rust or similar languages, with LLMs making large rewrites cheaper than perpetual emergency patching.
In that Galaxy Brain conversation, as OTB quotes it, Stamos gives an operational example from Cognition's work: rewriting a Mercedes-Benz application in roughly eight days with agent assistance versus months of human effort. Whether every utility or hospital stack can move that fast is an open question, but the direction is clear—proactive architecture, not patch Tuesday alone.
Stamos also critiques over-focusing on the power grid in social media panic while community hospitals and midsize vendors with one security person remain soft targets. His Decoder comments on isolated industrial networks still failing against physically infiltrative adversaries connect to our broader data-center and infrastructure protests coverage: resilience is not only cloud AI labs but every dependency chain.
What practitioners should do now
Stamos closes the published transcript sections with personal hygiene—unique passwords, password managers, passkeys, and skepticism toward AI-driven scams—that still matter when kill chains automate. For builders and security engineers, the interview implies a harder to-do list:
- Treat API keys at AI vendors as tier-zero secrets. The Fireworks-to-Cursor chain is a template; rotate keys and scope them as if they were production root credentials.
- Assume open-source offensive frameworks are commoditized. ARTEX is one name; forked agent frameworks will multiply. Baseline detections belong in SOC playbooks, not in quarterly pen tests alone.
- Price defender automation. If your investigation stack bills frontier-model prices per alert, model the cost against attacker economics Stamos describes.
- Plan memory-safe migrations on critical paths, especially internet-facing services still in C and C++.
- Read primary incident research. Check Point's ARTEX operator study and vendor advisories beat headline churn.
For teams running coding agents in production, pair those steps with guardrail products—Stamos's employer bought Corridor for Devin Security—and with external monitoring. When agent swarms can chain credentials across SaaS boundaries, use a dedicated control plane: AgentBeam's November attack timeline tracks the same incident cluster Stamos references and links defensive patterns for agent-written code paths.
What developers are saying
Speech bubbles clustered around one highlighted thread, standing in for HN debate on trust and AI content
Decoder had no Hacker News thread tied to this episode at publication time. On the separate Kevin Boone essay thread about AI slop and trust—another November 7 culture story—commenters wrestle with overlapping themes Stamos touches at the end of Decoder (misinformation and slop on social networks). User Avicebron argues that trusting versus distrusting content is the right distinction, but that giving up on truth because everything might be fake is what people lament (comment 50404962). That skepticism-first mindset is exactly what Stamos prescribes for reading security panic on X, even as he urges proactive rebuilding of real systems.
Bottom line
Stamos is not predicting permanent collapse. He is saying the assumptions the security industry grew on since the early 2000s no longer hold now that open-weight models automate kill chains, breaches chain through AI supply chains, and defender investigations can cost more than attackers spend in a month. The Verge interview is unusually concrete about names, economics, and SOC architecture. Treat it as a field report from someone who has led security at both a consumer platform and an AI software company—and plan for a dark age measured in years, not quarters.
Details reflect The Verge Decoder transcript dated November 7, 2026, and related public research cited inline.
Related reading
- Check Point reads ARTEX operator logs
- Chainalysis breach and agentic techniques
- OpenAI and Hacktron red-team partnership
- Cognition acquires Corridor for agent-written code guardrails
- Greenlight Luna autonomous SOC on Bedrock AgentCore
- DeepMind agents and RSA weaknesses — claimed vs verified
- Censys ARTEX exposed consoles count
- Vals AI CyberOffense: open-weight DeepSeek near gated GPT-6 Sol Daybreak — November 2026 update adds UK AISI four-to-seven-month gap context