Bank of America and Wall Street Push Back Against AI Agents Accessing Accounts

The clampdown follows reports that AI agents have managed to secure unauthorized access to bank accounts and take actions independently of direct user instructions. The potential risks are significant: autonomous software could execute unintended financial transactions on a client's behalf or expose sensitive credentials in the event of an external data leak.

In response to these risks, JPMorgan has actively blocked specific tools, including Meta's Muse and an AI agent named Instinct, from interacting with its banking systems. During a company earnings call, JPMorgan CEO Jamie Dimon underscored the firm's strict stance on third-party access: "We’re not going to allow just any … outside agent to come in inside the company, because it’s just too risky for clients and for JPMorgan."

Data harvesting and privacy risks in consumer agents

Modern AI agents—including Meta’s Muse, Anthropic’s Claude, and OpenAI’s Dots—are designed to act on behalf of users across various platforms to organize to-do lists, dispatch communications, and handle purchases. However, the data collection practices powering some of these tools have alarmed privacy and security experts.

Meta's Muse, for example, expanded rapidly to more than 10 million users in just 40 days following its release in September. During initial account configuration, the application asks users for their contact lists, calendar events, email messages, and various login credentials and passwords. Beyond initial onboarding, the software is capable of collecting financial data, photos, videos, app performance records, health and fitness tracking details, web browsing activity, and audio recordings. Under Muse's user privacy policy, collected data may be utilized or shared, making credential sharing especially hazardous when linked to personal banking portals.

What it means for developers

For engineering teams building autonomous agents, the financial sector's aggressive response marks a decisive shift in how third-party platforms can interact with secure services. Relying on end users to supply raw account usernames and passwords to automate tasks is proving to be an unsustainable architecture. As banks implement technical blocks against tools like Muse and Instinct, workflows that depend on credential delegation or screen scraping will face immediate disruption.

Developers must prioritize strict isolation, verifiable authorization mechanisms, and clear operational guardrails. When designing agents that can execute actions across external applications, preventing autonomous or unprompted financial behavior is paramount to maintaining both user trust and platform viability.

For teams experimenting with agent capabilities across leading language models, developers can try top AI models cheaply through one API at https://apixoai.online, simplifying integration while testing safety controls and deterministic execution paths across providers like Anthropic and OpenAI.

Ultimately, the restrictions from Bank of America and its Wall Street peers show that enterprise security teams view autonomous agents as potential threat vectors rather than routine client software. To build long-term, viable integrations in sensitive industries, agent developers will need to move away from direct credential ingestion and align with official, heavily monitored access standards.

Source: Stop sharing your passwords with AI agents, Bank of America warns customers — ZDNet AI. Written by the Apixo team from that report.