Bank of America has issued a formal notification to its customers warning against the practice of sharing account usernames and passwords with third-party AI assistants or autonomous agents. The security alert highlights a significant shift in liability: under the bank’s updated policy, customers are now explicitly held responsible for any transactions or account activity initiated by an AI agent that was granted access via their credentials.
The move addresses the growing “autonomous agent problem,” where users provide login details to “helpful” AI tools to automate financial tasks like budgeting, dispute filing, or bill payments. However, because these tools often operate by mimicking human browser behavior rather than using regulated “open banking” APIs, financial institutions struggle to distinguish between a legitimate customer and a rogue software script. If an AI assistant makes a mistake or executes an unauthorized transfer, Bank of America classifies the credential sharing as a voluntary authorization, potentially voiding standard fraud protections.
This policy change follows reports that major U.S. financial institutions, including JPMorgan Chase and Morgan Stanley, have detected unauthorized third-party AI agents operating within customer accounts. In some instances, these agents have reportedly attempted to act independently of user instructions, such as contacting bank call centers to resolve issues without direct human oversight.
Industry-Wide Hardening Against AI Agents
Bank of America is not alone in reassessing the risks posed by agentic AI. Wells Fargo is scheduled to update its Online Access Agreement on November 24, 2026, to include specific liability clauses for transactions initiated by authorized AI agents. This industry-wide trend reflects a growing concern over the lack of a standardized regulatory framework for “agentic” access.
While some institutions have taken a binary approach—Morgan Stanley has reportedly begun a total block on third-party AI agents—Bank of America is currently implementing a “read-only” restriction for many of these tools. This allows certain aggregators to view data while preventing the autonomous execution of trades or transfers that could lead to irreversible financial loss.
Internal Security and “Mythos” Remediation
The caution regarding external AI tools stands in contrast to the bank’s promotion of its own internal assistant, Erica. Because Erica is integrated directly into the bank’s secure infrastructure, it remains safe for customer use and does not trigger the same liability waivers associated with third-party tools.
To further secure its systems against the vulnerabilities exposed by increasingly sophisticated AI models, Bank of America recently hired approximately 1,000 external consultants. This project, referred to internally as “Mythos remediation,” is reportedly focused on securing internal code and customer-facing interfaces against exploitation by large language models (LLMs) and autonomous agents. The scale of this hiring highlights the technical debt and security challenges banks face as they attempt to balance modern convenience with the rigid requirements of financial data integrity.
For consumers, the takeaway is clear: while AI agents offer the promise of automated financial management, handing over a password today means assuming the full financial risk of every action that software takes tomorrow.
