While artificial intelligence agents are increasingly being used for tasks such as shopping and sending emails, early adopters are taking the technology a step further by giving these agents access to their bank accounts.
According to Axios, major U.S. banks have detected unauthorized third-party AI agents operating within customer accounts. In some cases, these agents appear to be acting on behalf of users who have provided their login credentials.
This creates a challenge for the traditional banking model. Even when customers authorize an agent to act on their behalf, bank policies or terms of service may require account holders to log in themselves.
The risks extend beyond questions of account access. AI agents with the ability to make purchases or transfer funds could create additional security concerns. Axios reported that some banks have monitored activity suggesting AI agents took actions independently of user instructions.
“This comes back to identity verification, and why traditional ‘authentication’ is not enough,” said Tracy Goldberg, Director of Cybersecurity at Javelin Strategy & Research. “FIs have to approach identity verification from a cybersecurity mindset, by establishing a consistent, risk-based digital identity standard. FIs should develop a baseline definition of digital identity that aligns with National Institute of Standards and Technology identity and authenticator assurance levels for onboarding, authentication/IDV, and transaction risk controls. Consumers associated with higher-risk environments may require additional document validation, device analysis, liveness testing, and/or step-up authentication.”
Highlighting Grey Areas
These developments raise several unresolved questions for banks. Distinguishing between human users, authorized AI agents, malicious bots, and cybercriminals is a significant challenge.
Even when an agent is authorized by a customer, banks have limited visibility into whether its actions align with the user’s intentions. These concerns are exacerbated by high-profile cases that have raised questions about the risks posed by AI systems, including those developed by OpenAI and Anthropic.
“FIs also must continuously evaluate identity trust using cyberthreat intelligence, which could identity rogue AI agents, if multiple accounts have been impacted,” Goldberg said. “Identity verification should extend beyond account opening. Institutions must continually reassess whether an identity can be trusted by incorporating cyber-threat intelligence.”
Considering a Compromise
For many banks, the security risks of agentic AI remain difficult to reconcile with its potential benefits. As a result, some financial institutions have moved to restrict third-party AI agent access. Similar concerns have emerged in retail, as evidenced by Amazon’s recent decision to block Meta’s Muse AI agent.
Despite these restrictions, consumer interest in agentic platforms remains strong. In its first 40 days on the market, Muse was downloaded more than 10 million times in the U.S., and the platform has demonstrated strong retention rates.
Recognizing the appeal of these tools, some banks are exploring a middle ground—read-only access. This would allow AI agents to view customer financial information, without authorizing them to initiate transactions or move money.
