How AI broke cybersecurity
Former Facebook and Yahoo security chief Alex Stamos explains why AI-driven hacking is overwhelming defenders, with open-source tools letting attackers find and exploit bugs in real time. He argues the industry is in for a rough adjustment period before things stabilize.
- Attackers now use AI to find vulnerabilities and write exploits in real time, inverting the old kill chain where research was done ahead of time.
- A chain of breaches at AI-related companies, including Fireworks and Cursor, let stolen API keys spread a bad plugin to about 5,000 companies.
- An open-source Chinese tool called ARTEX, paired with an uncensored open-weight model, has reportedly hit about 200 companies after first breaching three South Korean banks.
- Stamos compares the moment to the 1990s Wild West, but with the whole world now online, so the stakes are far higher.
Alex Stamos, chief security officer at Cognition and a longtime security researcher, argues that AI has broken the cybersecurity industry as it has operated for decades. The clearest sign, in his telling, is a run of breaches over the past month that hit AI-adjacent companies in sequence. Fireworks, a prominent provider of AI services, was reportedly breached first, and the stolen API keys were used to get into Cursor, the AI coding tool. From there, attackers pushed a malicious plugin that reached about 5,000 companies. Stamos also points to the Binance hack, which he calls the largest heist in human history, and to a wave of attacks on South Korean banks that has since spread to roughly 200 more organizations. He says he and his colleagues have been running at what amounts to a permanent high alert, with one CISO friend joking that everyone is at "DEFCON 2" through October and into November.
The mechanism behind this, Stamos says, is a reversal of the traditional attack pattern. Historically, attackers did the slow work ahead of time: finding a vulnerability, writing an exploit, then deploying it at scale. Now AI systems can find bugs and write exploits in real time, which compresses the whole "kill chain," the sequence of steps needed to achieve an effect on a target, into something close to automated. He names ARTEX, a Chinese open-source tool anyone can download from GitHub, as one example, and notes that when paired with a modified open-weight model it can break into a worrying number of organizations. The actors are a mix of state groups, including Russia and North Korea, and private criminals, many operating from places beyond the reach of law enforcement. Because they can use open-weight models instead of frontier systems with logging and guardrails, there is little visibility into what they are doing.
Stamos draws an analogy to the late 1990s and early 2000s, when Windows was a security disaster and a few hundred hackers at DEFCON could expose vulnerabilities that affected everyone. The difference now is scale: the internet runs the economy, so the same kind of exploit can do far more damage. He says the industry as currently structured is overdue for disruption and that what comes next will be a rough transition, a "dark age" in his words, while defenders work out how to adapt. He is not pessimistic about the long run, but he is blunt that the near term will be painful for companies that have not yet reckoned with automated attacks.