Alex Stamos Says AI Has Pushed Cybersecurity Into a Dark Age: the Fireworks-to-Cursor Breach Chain Hit 5,000 Companies

The cybersecurity industry has spent decades selling the idea that attacks can be contained. Alex Stamos, chief security officer at AI startup Cognition and formerly Facebook's security chief, thinks that premise just broke. In an interview on the Verge's Decoder podcast published Friday, he described the past month as a turning point where AI-boosted attackers began outpacing the entire defensive industry at once.

His accounting of October is bleak by any standard. A series of severe bugs in foundational software — PHP and Node.js among them — was followed by a run of breaches at AI "neocloud" providers, the Binance hack that Stamos called the largest heist in human history, and the Cursor incident. "We have been at DEFCON 2 for all of October and now going into November," he said, quoting a CISO friend asking whether anyone else was tired.

The most instructive part is how the incidents chained together. According to Stamos, one of the breached companies was Fireworks, a prominent provider of AI inference services; API keys stolen from Fireworks let attackers break into Cursor, the AI coding tool; once inside Cursor, they pushed a malicious plugin that reached roughly 5,000 more companies. One initial compromise became thousands of potential downstream breaches — precisely the propagation pattern AI-accelerated operations make cheaper to execute.

Stamos attributes the shift to adversary capability, not just volume. Open-weight models have reached the point where, in his view, they can automate most of the "kill chain" — the sequence of finding a vulnerability, writing an exploit and turning access into money or leverage. He pointed to a mix of criminal groups and state actors, including Russia and North Korea, now breaking into systems at a pace defenders cannot match with manual processes.

His forecast is not unconditional doom. Stamos compared the current moment to the 1990s, when the first always-online generation of the internet produced a decade of chaos before the security industry matured. The problem now, he argues, is that the defensive playbook — perimeter tools, awareness training, quarterly audits — was built for human-speed attackers, and rebuilding it for machine-speed ones will take years of uncomfortable adaptation.

The interview lands in a week when AI security is no longer theoretical. Just this week, researchers documented hundreds of exposed consoles for Artex, an AI-driven hacking framework tied to real attacks, while DOJ prosecutors charged a $129 million impersonation scheme built on AI voice and video clones. The defensive side is responding in kind — AI agents now find and disclose vulnerabilities like the recent OpenSSH bug — but Stamos's core point stands: for the first time, both sides of the industry are running on the same models, and the attackers got there first.

For companies buying AI tooling, the practical takeaway is uncomfortable: your vendor's vendor is now your attack surface. The Fireworks-to-Cursor chain showed that a single API key leak in the AI supply chain can metastasize into thousands of incidents before any single victim knows it happened. In Stamos's dark age, the scarce resource is not intelligence — it is trust that can be verified at machine speed.