AI is reshaping cyberattacks faster than defenses can keep up

In an interview with The Verge’s Decoder, Stamos described a run of breaches involving AI-related companies and other targets. He said stolen API keys from Fireworks were used to access Cursor, where attackers pushed a malicious plugin to 5,000 companies, creating the potential for further breaches. He also cited attacks on South Korean banks and about 200 other companies, and said the Binance theft was carried out by North Koreans using human intelligence and social engineering, with AI’s role unclear.

Stamos said open-weight models are lowering the expertise needed to carry out attacks. Tools can now search targets, identify software flaws and generate exploits as they go, rather than relying on specialists to prepare attacks in advance. He pointed to ARTEX, a Chinese open-source tool, as one example. The change, he argued, weakens the old division between highly skilled researchers and less-skilled operators, as well as the economic and professional norms that helped shape cybersecurity in the 2000s and 2010s.

The pressure is not limited to software bugs. Stamos said security teams must also use AI to analyze the huge volumes of logs generated by companies and help investigate suspicious activity. He described a layered approach: simpler classifiers scan data first, then smaller and larger language models can be called as needed. But he warned that powerful investigations can be costly, citing a colleague’s estimate of $900 for one investigation using a frontier cyber model. Defenses must become cheaper if they are to compete with attacks whose marginal cost is approaching zero.

Stamos urged companies to test their own systems with AI red-team tools or outside specialists, while warning that these products are new and can cause disruption. Defensive agents, he said, should have firm limits on what they can do without human approval. He also argued that some legacy software may be cheaper to rebuild than to keep patching, proposing that Google and Amazon consider a memory-safe Linux kernel for production servers. He expects a difficult period rather than permanent catastrophe, and said signs of improvement would include fewer major breaches and fewer serious vulnerabilities in core software.