Bank executives are rushing to restrict access to their systems following reports by Axios that AI is advancing faster than financial institutions can react. In a post on Substack, Gary Marcus contextualizes these reports within a series of recent incidents and calls for regulation not to focus solely on the most advanced models.
Axios journalist Madison Mills writes that these findings have alarmed bank executives, who are seeking to limit access to their systems. Gary Marcus also cites other episodes. On October 11, attacks on supermarket refrigeration systems in France may have been facilitated by agents, meaning AI systems capable of acting autonomously to complete tasks. A wave of cyberattacks is also hitting Japan. Meanwhile, The Information has mentioned a Latin American hacker group, Cybercartel, though the source did not elaborate on its activities.
For Gary Marcus, these events show that malicious actors do not need so-called ‘cutting-edge’ models to cause damage. He argues that even if the development of these models were to stop, the misuse of existing models would pose serious problems. He cites various domains, from education to politics, with deepfakes and large-scale disinformation, and banking. Ignoring these risks in favor of speculative scenarios would, in his view, be a mistake.
A diagnosis made as early as 2025
Gary Marcus recalls that he warned, in early 2025 in Politico, that cybersecurity was facing a ‘perfect storm’ for four reasons. First, generative AI, whose popularity is growing while its price is falling, is an ideal tool for attackers, capable of producing credible texts, like phishing emails, and fake videos at low cost. He cites the case of a cybercriminal team that defrauded a Hong Kong bank of $25 million.
He then points to the vulnerability of large language models to hijacking, such as prompt injection attacks, for which, he says, there is no known solution. He also notes that automatically generated code has sometimes introduced new security flaws, with some developers not fully understanding the produced code. Finally, he mentions the deregulation policy of the new US administration, which could weaken investigation and oversight, risking, in his view, leaving the world ‘quite vulnerable’ to increasingly bold attacks.
Gary Marcus makes three predictions: the AI industry will not take real responsibility, regardless of the damage caused by generative AI; the Trump administration will do little to hold it accountable; and ‘things are going to get worse.’
