Banks and AI Firms Prepare for Cyber Risks Amid Agentic AI Concerns

America's financial institutions have joined the largest AI companies in planning for potential disasters. Such preparations signal the scale of the risks that major corporations perceive. Concerns about frontier AI systems, rather than AI agents, have prompted banks to invest in prevention and prepare for worst-case outcomes.

Banks are encountering unvetted AI agent activity, mainly because customers give agents login credentials, as reported on Thursday. Some agents appear to have acted without instructions from users. Banks fear the wider consequences of customers allowing agents to access accounts and perform actions that customers may not authorize. This activity differs from a summer incident in which an AI agent swarm hacked a platform at the startup Hugging Face, triggering an industry security crisis.

However, banks say banks have detected neither rogue agent activity nor evidence that a hack affected customer accounts or funds. Banks strongly discourage customers from giving passwords to agents. Services already available or under development can provide read-only access. Plaid is one platform that can retrieve account balances and transaction histories, among other functions.

A JPMorgan spokesperson said customers must never share passwords with any person or thing, citing secure open banking alternatives and customers' responsibility for account protection. JPMorgan and other banks have acted to block agent access. However, JPMorgan is investigating secure methods for customers to connect accounts to services customers choose.

Major banks participated in early tests of Anthropic's Mythos, the AI lab's most capable models for cyber tasks. AI labs and major banks used the testing to improve understanding of cyberattacks and other risks and prepare for those threats. Banks have since developed plans to prevent AI-related problems and to safeguard accounts, reassure customers and maintain operations during a large-scale attack. Critical infrastructure providers, including financial institutions, are required to undertake scenario planning to meet responsibilities to government, shareholders and customers.

JPMorgan CEO Jamie Dimon warned in Chicago earlier this month about cyberattacks and the possibility that AI agents could act without users' approval. Dimon said vigilance and management were necessary because stopping AI would not be possible. An economist last month raised the possibility that agents could trigger a bank run by transferring users' household cash into accounts offering higher yields.

The banking concerns coincide with growing fears of a cybersecurity crisis. Security researchers found major vulnerabilities in a programming language that supports millions of websites, representing one of AI's most consequential cybersecurity discoveries so far. An Anthropic security employee warned, "as bad as things have gotten in the past 4 days, they are going to get a lot worse before they get better." Salesforce CEO Marc Benioff quoted that warning and posted "code red".

The Thursday reporting about banks concerned agentic AI rather than frontier AI systems. The disclosures prompted demands for regulation across political divisions, including from Senator Bernie Sanders, Florida Governor Ron DeSantis and Senator Mark Warner. Former White House AI czar David Sacks said banks should accommodate user-directed agents with access users can grant, restrict or withdraw, initially through read-only access. Sacks also called it irresponsible to equate user-authorized agent activity with cyberattacks.

AI companies expect a political backlash, while banks are preparing for possible operational disruptions. The extent of federal preparations for a crisis affecting both sectors remains unclear. Rapid AI progress is pushing Wall Street and Silicon Valley to prepare for worst-case outcomes while Washington seeks to reduce cyber risks and relies on voluntary safety commitments.