Citi, PNC and Citizens tell customers they own any loss an AI agent causes Citibank updated its Online User Agreement this week to make customers responsible for any loss caused by third-party AI agents they grant account access to, following PNC's November 18 terms change and a similar warning Citizens Bank emailed customers this week; JPMorgan Chase and Bank of America moved in October. Citi said it wanted to "provide greater clarity about those risks and to reinforce that customers who choose to grant account access to third-party AI agents are responsible for activities conducted through those agents," and its terms also place the loss on customers if an AI tool leaks credentials or is hacked. Wells Fargo's Online Access Agreement, effective November 24, counts every transaction by a customer's delegates, AI agents included, as authorized "including those you did not want or intend," a reading that removes agent transactions from Regulation E's capped-liability protections — and no court or US regulator has yet tested it. If you hand your banking login to an AI assistant and it gets tricked or breached, your bank's terms now say the loss is yours. You'll want to read the agreement before you connect anything. Citibank updated its Online User Agreement this week to say customers who give third-party AI agents access to their accounts are responsible for what those agents do, according to TechCrunch. PNC changed its terms on Wednesday, November 18, and Citizens Bank emailed customers the same warning this week. JPMorgan Chase and Bank of America had already moved in October. Citi's own explanation, quoted by TechCrunch, is that it wanted to "provide greater clarity about those risks and to reinforce that customers who choose to grant account access to third-party AI agents are responsible for activities conducted through those agents." That's the polite version. The sharper detail is that Citi's terms also put the loss on you if an AI tool leaks your credentials or is hacked, and a criminal then uses them. The answer to "are banks liable if my AI agent gets hacked" depends on one word. Under Regulation E, which implements the Electronic Fund Transfer Act, a consumer's liability for an unauthorized transfer is capped, and the cap is low if you report quickly. Banks that define everything an agent does as authorized by you take those transactions out of that category before a dispute starts. Wells Fargo's Online Access Agreement, effective November 24, says every transaction by a customer's delegates, AI agents included, counts as authorized, "including those you did not want or intend." JPMorgan's October terms likewise put the consequences of giving an agent your credentials on the customer. No court or regulator appears to have tested this reading, and no US agency has publicly weighed in, so whether those clauses survive a real dispute over a drained account is open. Frankly, a bank deciding that a stolen password is a customer instruction is a big claim to make in a contract. JPMorgan tests AI agents that trade currency with no human in the loop https://startupfortune.com/jpmorgan-tests-ai-agents-that-trade-currency-with-no-human-in-the-loop/ JPMorgan and Hyundai Card ran a simulated test in which Hyundai's agent asked for euro top-ups and JPMorgan's Kinexys unit executed them within preset limits. A multi-agent follow-up is planned for the first half of 2027, with no commercial launch date. - JPMorgan AI agents trade currency autonomously https://startupfortune.com/jpmorgan-tests-ai-agents-that-trade-currency-with-no-human-in-the-loop/ - Hyundai Card automated foreign exchange test https://startupfortune.com/jpmorgan-tests-ai-agents-that-trade-currency-with-no-human-in-the-loop/ JPMorgan's October move was more blunt. CNBC reported on October 13 that Chase is blocking outside agents such as Muse and Instinct from reaching accounts directly. Bank of America, for its part, asked customers on October 27 not to share their banking password with an AI assistant, TechCrunch reported at the time. What an agent can actually get wrong The risk isn't hypothetical hand-wringing. A browser agent that logs in for you can read a malicious web page, and a prompt hidden in that page can tell it to send money somewhere else. A shopping agent can buy the wrong thing. A finance assistant that stores your username and password becomes a single place where all of it can leak. The bank can't tell from the login whether you or your software is on the other end, and the new terms settle that doubt in the bank's favor. Anthropic's recently launched personal finance tool takes the opposite route, according to TechCrunch. It's marketed as read-only, with no ability to "move, manage, or invest your money." That's the design the banks seem to prefer. The squeeze on agent-payment startups If you're building an agent that pays bills or moves money, this is a distribution problem as much as a legal one. Your product depends on a bank letting your users' credentials through, and now the bank's contract tells those users that you are the risk they own. Few customers will accept that for a convenience feature. Startups that can offer scoped, revocable access through an open-banking API, with no passwords handed over, are in a stronger position than ones that log in as the user. For the bigger banks, the incentive is plain. They collect nothing extra for the risk, and the agents' makers carry no obligation to them. Passing the loss to the account holder costs them a paragraph of legal text. Wells Fargo's new terms take effect on November 24, which is the next date to watch, along with whether any regulator weighs in on how Regulation E treats an agent's actions. Until then, the practical rule from these banks is the same: if you let an agent in, you're signing for everything it does. Also read: How to handle LLM API rate limits in production before retries burn your quota https://startupfortune.com/how-to-handle-llm-api-rate-limits-in-production-before-retries-burn-your-quota/ • FBI and Google warn AI hacking tools like Artex now break into networks in hours https://startupfortune.com/fbi-and-google-warn-ai-hacking-tools-like-artex-now-break-into-networks-in-hours/ • Oracle and OpenJDK patch Java4Shell, a critical Java flaw found with AI https://startupfortune.com/oracle-and-openjdk-patch-java4shell-a-critical-java-flaw-found-with-ai/ Goldman Sachs pulls Claude Code and GitHub tools from traders and bankers https://startupfortune.com/goldman-sachs-pulls-claude-code-and-github-tools-from-traders-and-bankers/ About 12,000 Goldman Sachs engineers keep full access to Claude Code and GitHub's coding tools, as do "strats" who blend trading and coding. The October guidance change follows CIO Marco Argenti's July praise of "vibe coding" among non-technical staff. - Goldman Sachs restricts Claude Code for traders https://startupfortune.com/goldman-sachs-pulls-claude-code-and-github-tools-from-traders-and-bankers/ - AI coding tools banned from banking staff https://startupfortune.com/goldman-sachs-pulls-claude-code-and-github-tools-from-traders-and-bankers/ This article is posted in AI News https://startupfortune.com/category/ai/ , check it out for more related stories. Join the discussion Open in the community → https://startupfortune.com/community/ Almost there. Sign in and your reply posts straight away.