What is happening
In an interview on The Verge's Decoder podcast, Alex Stamos, chief security officer at Cognition and a lecturer at the Stanford Cyber Policy Center, describes an October ironically dubbed 'security awareness month' and marked by a string of breaches. According to the source, several AI-related companies and so-called 'neoclouds' were compromised in succession, along with incidents cited such as the Binance hack, which Stamos called the largest theft in human history, and the Cursor incident.
Stamos reconstructs a chain of attacks: from the compromise of Fireworks, identified as an AI services provider, API keys were reportedly stolen and used to break into Cursor; from there, a malicious plugin was reportedly spread to about 5,000 other companies, with as many potential breaches. This is the phenomenon the expert calls 'continuous permanent breach' in the technology sector. A fellow CISO, he reports, asked in a group whether others were tired: the response was unanimous, because the sector has effectively been at maximum alert level throughout October and now November as well.
The attack chain reverses
The central point of the analysis concerns automation. Stamos argues that open-weight models have reached a level at which they can automate much of the 'kill chain,' the sequence of steps that leads to an effect on a target: stealing information, money or bitcoin, or shutting down systems. In the traditional model, he explains, vulnerability research and exploit writing happen in advance, and then a human distributes them against one or many targets. With AI, according to Stamos, the process is reversed: a system examines targets, finds bugs and writes exploits in real time.
As an example, he cites ARTEX, described as a Chinese open-source tool available on GitHub, including in English translation, which, if powered by models such as an 'obliterated' GLM-5.3, would make it possible to breach a surprising number of organizations. The attack reportedly began with three South Korean banks and then hit about 200 other companies. Stamos adds that he helped some companies fix the vulnerabilities, clarifying that Cognition is not an incident response company. The historical comparison is with the 1990s: then, he recounts, a DEFCON conference brought together a few hundred people and many were benevolent researchers, while today the ecosystem is much broader, with economic incentives and actors beyond the reach of law enforcement.
State actors and the 'dark age'
Asked who is behind the attacks, Stamos responds that they involve different groups: not only private actors, but also states such as Russia and North Korea, which according to the source use AI to identify and exploit vulnerabilities. The expert also highlights a technical and economic aspect: open-weight models can be run locally, without the controls and logs available to law enforcement on frontier models; he cites as an example using a quantized and 'obliterated' GLM-5.3 on one's own computer.
Stamos does not describe himself as a long-term catastrophist, but he believes that we have entered a 'dark age' of adaptation to new capabilities. The interview, the first episode of a special series on AI in real life, also touches on how disinformation and low-quality AI-generated content on social networks are managed, without providing operational guidance. It therefore remains to be verified what the scale and attribution of the cited incidents are, the actual number of companies involved in the Fireworks-Cursor chain, the scale of the Binance hack and the effectiveness of the defense tools mentioned: elements that the source reports as the interviewee's statements, not as independently verified data.
