Breaking news: U.S. banks are taking hits from AI agents

Gary Marcus used his newsletter to amplify an Axios report by Madison Mills finding that U.S. banks are taking hits from AI agents, with findings that alarmed banking executives who are racing to restrict access to their systems. In Marcus's framing, AI is moving faster than financial institutions can respond, a situation he describes as a set of day-after scenarios already arriving. He pairs the news with a quoted remark from Sam Altman dated October 3, in which Altman said the world should accept some bad things happening for the benefits of the technology and people having agency. The briefing does not detail which banks were affected, how the agents were used, or what specific losses occurred.
Marcus places the banking report in a wider cluster of incidents. He points to attacks on refrigeration systems in France on October 11 that may have been facilitated by agents, a growing and ongoing wave of cyberattacks in Japan, and months of attacks he attributes to OpenAI and others against targets including Hugging Face and the Australian government. He also cites reporting from The Information the previous day on a Latin American hacking group called Cybercartel. The through-line he draws is that malicious use of agents is escalating across unrelated sectors, from finance to physical infrastructure.
The argument Marcus makes about frontier models is the analytical core of the piece. He contends that hackers do not need state-of-the-art frontier models to cause mayhem, so regulatory attention fixed on the largest, most compute-intensive systems would miss the real problem. Even a halt to future frontier development, he writes, would leave serious harms from existing models. He calls for regulation that addresses current models and present-day misuse across domains including education, politics through deepfakes and disinformation, and banking, rather than focusing only on speculative future capability.
He recalls a Politico piece from early 2025 in which he warned of a perfect storm in cybersecurity, built on four factors. The first is that generative AI is cheap, popular and effective for attackers: despite unreliability and hallucinations, it produces plausible text for phishing and deepfaked video at almost no cost, widening the pool of targets. He cites a cybercrew that bilked a Hong Kong bank out of $25 million. The second is that large language models remain susceptible to jailbreaking and prompt injection, for which he says no known solution exists. Prompt injection, in practice, is when text fed to a model, such as the contents of a web page or document, contains instructions that hijack the model's behavior, a hazard that grows as agents are given tools and the ability to act. The third is that AI-generated code is increasingly used by developers who may not fully understand it, and such code has in some cases introduced new security holes. The fourth is a U.S. administration he describes as determined to deregulate, cut costs and publicly shame employees, which he expects to reduce the quality and quantity of enforcement and investigations.
Marcus closes with three predictions: that the AI industry will not take real responsibility no matter how much chaos generative AI causes, that the Trump administration will do little to hold it responsible, and that things will get worse. These are stated as forecasts rather than reported findings, and the incidents he lists, including the France refrigeration attacks, are described in hedged terms such as may have been facilitated by agents.
For developers and product teams, the practical reading is that agent security is now a live operational concern rather than a hypothetical one. This suggests teams building agents with tool access should treat prompt injection as an unsolved design constraint, limit what agents can reach, and log their actions, since a hijacked agent can act with the permissions it has been granted. A likely trade-off is that tighter restrictions on what agents may access will slow some useful automation. The banking response Marcus describes, restricting system access, is one version of that trade-off, and inference from the facts here is that other sectors may follow. Regulators, meanwhile, face a choice between policing only the largest models and addressing the cheaper, widely available ones already in use.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Who knows what piece of infrastructure might be attacked next.