Anthropic launches small-business agent as banks block AI assistants.

Some of the world’s largest banks are tapping the brakes on autonomous agents, especially in customer-facing finance. Citigroup and Bank of America confirmed on 19 October that they have barred third-party AI assistants from logging into retail banking accounts ([1]). These firms, along with JPMorgan Chase, Wells Fargo and others, say they are detecting unvetted bots entering customer accounts and even attempting “independent” actions beyond user instructions ([2]). The fear is that an unattended agent with access to personal finances could move money without a customer’s knowledge, for example by sweeping cash into higher-interest accounts and inadvertently undermining a bank’s deposit base. One analysis suggests that if AI agents automatically shifted enough funds to earn around 3% interest elsewhere, leading US banks would face up to $47 billion in extra annual interest costs ([3]).

Bank of America has said it will maintain its ban on outside AI agents until liability for AI-initiated payments is clearly defined ([4]). The banks are working with fintech partners on safer models that grant bots limited, read-only data access under existing open banking rules ([5]) ([6]). Executives and regulators have signalled that truly autonomous transactions, where an AI moves money or makes payments unchecked, are off the table unless and until customers and providers have reliable ways to authenticate agents and assign responsibility for their mistakes ([7]) ([8]). The upshot for business leaders is that even as AI agents become more capable, they may face interim constraints in high-stakes areas like finance. That puts a premium on strong internal controls. In fact, a recent Deloitte survey of CFOs found that 74% would need days or weeks to pause an AI application if costs ran over budget, and only 7% have real-time kill switches in place ([9]) ([10]). The push by banks to preempt “rogue” agents is a reminder that organisations must set boundaries for AI systems, from spend limits to human sign-offs, before granting them authority over financial transactions or sensitive data.

A case unfolding in South Korea has underscored the security stakes of autonomous AI. On 26 October the Seoul Metropolitan Police Agency announced the arrest of seven people accused of using an AI-driven hacking tool to breach more than 90 companies and steal some 8.2 million customers’ and employees’ personal records ([1]). The group’s leader, a former cybersecurity professional, allegedly deployed an AI penetration-testing agent to scan for vulnerabilities and carry out attacks at a scale and speed that would be hard for any human team to match ([2]). Investigators say the suspects began using the AI tool in August to dramatically expand their campaign, which ranged across hospitals, tech firms, schools and even banks ([3]). This was the first major arrest linked to AI-assisted hacking in the country, with police warning that advanced AI has “lowered the barrier” for criminals lacking traditional hacking skills ([4]).

The incident in Korea comes on the heels of another AI-enabled hacking spree observed earlier in October, when a Chinese-speaking hacker used an open-source agent platform to break into more than 100 companies’ websites on his own ([5]) ([6]). Together, these examples demonstrate how autonomous tools can be exploited to quickly find and exploit security gaps without human supervision. It’s a stark illustration of why regulators and enterprises have been raising flags about rogue AI agents ([7]) ([8]). For corporate security and risk officers, the takeaway is the importance of robust safeguards, from monitoring AI-driven activities to restricting critical system privileges, to defend against both malicious bots and well-meaning agents that might go off-script.