Alex Stamos, Chief Security Officer (CISO) at Cognition and a lecturer at Stanford University, believes cybersecurity is going through a particularly chaotic period, with artificial intelligence amplifying waves of intrusions and attacks. In an interview with The Verge’s *Decoder* podcast, he states that the security industry as we know it is “essentially behind” these offensives and is entering what he calls a “dark age.”
- Alex Stamos affirme que l’IA bouleverse la cybersécurité et que le secteur accuse un retard face aux attaquants.
- L’IA peut automatiser les attaques, de la détection des failles à l’écriture d’exploits en temps réel.
- Le coût des attaques tend vers zéro, alors que la défense reste très coûteuse.
- Il est plus fiable de prouver l’authenticité d’une image que de détecter les contenus générés par IA.
To illustrate this point, Alex Stamos cites several recent incidents. According to him, the company Fireworks, an AI service provider for several businesses, was hacked, and API keys (the credentials that allow software to communicate with each other) were allegedly stolen. These keys allegedly allowed access to Cursor, where a malicious plugin was reportedly distributed to 5,000 companies, potentially causing 5,000 breaches. He also describes the Binance hack as “the biggest theft in human history.”
He also describes how these attacks are carried out. He talks about a “kill chain,” meaning the series of steps required to reach a target. Traditionally, attackers would prepare vulnerabilities and exploit programs in advance, before a human deployed them. With AI, particularly open-weight models (whose workings are accessible to all), this chain could be automated: according to him, a system can spot vulnerabilities and write exploits in real time. He cites a Chinese open-source tool available on GitHub which, combined with a model named “obliterated GLM-5.3,” allegedly allowed intrusion into three South Korean banks and then, according to him, around 200 additional companies.
A World Resembling the 1990s
Alex Stamos compares the situation to the late 1990s, when Windows was, in his words, “a security disaster.” The difference, he specifies, is that the entire world now runs on the internet, and attacks can now be automated at large scale. He also discusses the actors involved: according to him, attacks on AI companies are carried out by what is called the “Chinese cottage access industry,” small Chinese subcontractors who resell their access to larger companies, while the bitcoin theft at Binance would primarily be the work of North Korea, involving, he says, very elaborate social engineering.
Defense Costs Are Too High
Economic imbalance is central to his argument. He contends that the cost per token for attackers is trending toward zero, while the cost of defense remains high. He cites a colleague working at a major security platform who says a single investigation conducted with a cutting-edge cyber model costs $900. To address this issue, Cognition trains its own models, or retrains open-weight models, to reduce these costs.
Alex Stamos also mentions a more radical solution: rewriting the Linux kernel in Rust, a memory-safe language. He estimates that such a project, led by Google or Amazon, would cost $100 million, a sum he considers modest compared to the societal costs of successive patches. He adds that the number of patches released each month continues to increase: according to him, the most recent Patch Tuesday edition had “900-something” patches.
Asked about content verification, he believes it is more reliable to prove the authenticity of an image than to detect AI-generated content. He notes that Adobe already offers cryptographic verification for professional photographers, and that iPhones now include a similar feature.
