US Presses Banks On AI Cyber Threat

US Treasury secretary Scott Bessent convened the chief executives of some of the country’s largest banks this week to examine the growing cyber risks posed by a new artificial intelligence model developed by Anthropic, according to people familiar with the discussions.

The meeting brought together senior figures from Bank of America, Citigroup, Goldman Sachs, Morgan Stanley and Wells Fargo, many of whom were already in Washington for a separate industry gathering. Jamie Dimon, chief executive of JPMorgan Chase, was invited but unable to attend.

Also present was Jay Powell, underlining the level of concern within US economic and regulatory circles about how rapidly advancing AI systems could reshape cyber risk across the financial system.

The focus of the discussion was Anthropic’s latest model, known as Claude Mythos Preview, which has drawn attention for its ability to identify software vulnerabilities at a level that, according to the company, exceeds that of most human experts. Officials are increasingly concerned that such capabilities, while potentially valuable for defensive purposes, could also be exploited by malicious actors.

The meeting highlights a shift in how policymakers are approaching AI risk. Rather than viewing the technology solely through the lens of productivity and growth, authorities are now grappling with its implications for systemic stability, particularly in sectors such as banking where cyber resilience is critical.

Warnings from within the industry are not new, but they are becoming more pointed. In his annual letter to shareholders this week, Dimon described cyber threats as one of the most significant risks facing the financial system, adding that artificial intelligence would almost certainly intensify the challenge. He noted that banks would need to commit substantial resources to strengthening their defences in response.

Anthropic’s decision to restrict access to its new model reflects similar concerns within the company itself. Unlike previous releases, Claude Mythos Preview has been made available only to a limited group of partners, including Amazon, Apple and Microsoft. The aim, according to the company, is to give trusted organisations an early opportunity to identify and mitigate vulnerabilities before broader deployment.

The model is described as general purpose, with applications beyond cyber security. However, its capacity to detect weaknesses in software systems has become the central issue. Anthropic said the system had already uncovered thousands of serious vulnerabilities, including flaws in major operating systems and web browsers that had gone undetected for years.

In a statement accompanying the release, the company said AI models had now reached a level of coding proficiency where they could outperform all but the most skilled human specialists in identifying and exploiting such weaknesses. That assessment has reinforced fears that the balance between defensive and offensive cyber capabilities may be shifting.

Anthropic has been in contact with US government officials in recent months to discuss both the risks and potential safeguards associated with its technology. Those conversations have centred on how to manage the dual-use nature of advanced AI systems, which can be deployed to strengthen security but also to undermine it.

The limited rollout of Mythos also follows a series of internal setbacks. The company recently disclosed two incidents in which sensitive data, including material related to the new model and elements of the underlying code for its Claude assistant, were leaked online. Anthropic attributed both breaches to human error, but they have nonetheless raised questions about its own security controls at a time when scrutiny is intensifying.

For policymakers, the episode underscores the broader challenge of governing AI development at a pace that matches technological change. Financial institutions, long accustomed to managing cyber threats from criminal groups and state actors, must now prepare for a landscape in which the tools available to attackers are becoming more sophisticated and widely accessible.

The involvement of the Federal Reserve signals that concerns extend beyond individual institutions to the stability of the financial system as a whole. A successful cyber attack on a major bank or piece of financial infrastructure could have cascading effects, particularly if enabled by more advanced AI-driven techniques.

Despite these risks, banks are also investing heavily in AI to improve their own operations, including fraud detection, customer service and internal security. This creates a complex dynamic in which the same technology that offers efficiency gains also introduces new vulnerabilities.

None of the institutions involved in the meeting publicly commented on the discussions. Anthropic declined to address the gathering directly, while representatives for the Treasury, the Federal Reserve and several of the banks either declined to comment or did not respond to requests.

What is clear is that the tone of the debate is changing. AI is no longer seen as a distant or abstract risk in financial services. Instead, it is emerging as an immediate operational concern that requires coordination between government and industry.

The challenge for regulators will be to strike a balance between enabling innovation and preventing misuse. For banks, the priority is more immediate: understanding how quickly these tools are evolving, and ensuring their systems can withstand threats that are becoming more complex by the month.

As AI capabilities continue to advance, the discussion initiated by Bessent and his counterparts is unlikely to be a one-off. It marks the beginning of a more sustained effort to assess how emerging technologies could reshape the risk landscape at the core of the global financial system.