Citi rewrote its online banking terms this November. Give an outside AI agent your Citi login, and Citi says it will treat what the agent does as authorized by you. If the agent leaks that login and someone else uses it, Citi says it will treat their actions as authorized, too. JPMorgan said in October it would block outside agents, and Bank of America asked customers not to share their banking password with an AI assistant. Citizens Bank emailed customers this week with a similar warning, and PNC split the question in two: a bank-approved connection can get limited access, but handing an agent a password or a one-time code is a different thing, and PNC's new agreement treats it that way.

What Citi's Terms Actually Decide

Strip the legal language and the decision underneath is simple: who gets to act inside a retail account, and what does the bank check before it lets that happen. Under Citi's terms for third-party agents, the shared login stands in for the customer's instructions. If an outside AI agent holds the customer's login, Section H of Citi's terms says the bank will consider the agent's actions authorized by the customer, will not investigate whether they conflict with the customer's instructions, and leaves monitoring the agent to the customer. Citi's own agents are carved out of the rule, which tells you something: the bank trusts the agents it built and can see inside, and it does not trust the ones a customer brought in from outside.

Who finally pays is a separate question, and a contract does not settle it alone. Under Regulation E, a transfer made by someone the customer gave access to is generally treated as authorized until the customer tells the bank otherwise. But the CFPB says transfers made with credentials obtained from a third party by fraud can be unauthorized transfers, and an account agreement cannot waive those protections. Citi's own terms hold the customer responsible for an agent that exceeds its authority "except as otherwise permitted under applicable law."

This isn't a theoretical worry. Axios reported on October 22 that banks were finding outside AI agents inside customer accounts, in some cases taking actions with no clear matching instruction from the customer. Many of those agents log in with the customer's own username and password, so at the login screen the bank cannot tell the customer from the agent. Federal Reserve Governor Michael Barr named the same gap in Milan on November 16: banks report agents attempting actions in retail accounts without being able to verify the agent's authenticity, the customer's consent, or whether the agent is following instructions. He called for strict authentication, explicit consent verification for specific actions, and technical controls that limit what an agent can do. The OCC has since issued AI examiner guidance touching outside agents, but it is nonbinding and creates no new requirement on its own.

The liability question moved faster than the control did.

Why a Login Is the Wrong Thing to Trust

A login is a bearer secret. Whoever presents it gets treated as the account holder, regardless of what they actually do once inside. That already fails the moment software is the one holding it, because software copies, logs and forwards what it holds.

The crypto side shows the same failure wearing different clothes. Lighter blamed a drain of roughly $10.4 million on a trader's leaked API key, a pattern we've looked at before in why AI agents keep getting their crypto stolen. The Safe wallet that lost $6.4 million is the more instructive case for banks: no owner signed anything. A role key moved the money through a delegated permission the owners had set up, and one analyst traced it to a trading policy that checked a trade's beneficiary but not who executed it. One is a leaked bearer token, the other a delegation scoped wider than anyone meant, but both answer the same wrong question: does the caller hold the secret or the role, instead of is this specific action allowed right now.

Last week we looked at JPMorgan blocking the card number instead of writing the spend rule. This month the banks wrote contract language assigning blame instead of a check that settles the question before the transfer happens.

The Condition a Bank Actually Needs to Ask

The fix isn't a better password or a longer one-time code. It's asking a narrower question: not does this caller have the login, but does this specific authority still hold right now. An ERC-7710 delegation check, offered on Base, answers part of that, as a signed yes or no. It comes back true only when the agent's wallet is the delegate, the delegator is the principal the bank expects, the principal's signature verifies, the delegation is unrevoked as of the block checked, and any time window is still open. The amount and target limits the principal signed come back beside the verdict as declaredLimits.

The signed yes proves the grant, not the payment. Say a customer delegates an agent to pay a utility up to $200 in a stablecoin, through Friday, and on Thursday the agent asks to send $350. The delegation check can still come back true: the grant is signed, unrevoked and inside its window. The result also carries the $200 limit, so the bank compares the $350 request against it, applies its own policy to the payee, and refuses. The attestation answers whether the authority is live; the bank still decides whether this transfer fits it.

A result that carries a delegation expires in five minutes, because revocation is one transaction away. Other results are good for up to 30 minutes. Account code conditions do something adjacent: they distinguish a plain key from an EIP-7702 delegation or a contract, without ever exposing what the code contains. Up to ten conditions can run in a single call to POST /v1/attest, which reads 37 chains in all, and the caller learns only whether the condition it asked about was met. Boolean, not balance, and never the login itself.

Hong Kong's Financial Secretary Paul Chan put the stakes in one sentence at the Bloomberg Global Regulatory Forum: "An algorithm has no license to lose and no reputation to protect. The institution has both." That's an argument for a specific kind of proof: the institution needs one fact it can point to after the money has moved, not the agent's full conversation log. The HKMA's Project SAGE is already examining agent identity and delegated authority, alongside the interoperability and liability questions Chan raised.

The Record That Outlives the Argument

Securities and Futures Commission CEO Julia Leung has set the bar her regulator will test firms against: can a firm reconstruct the decision, explain the outcome, and determine accountability, given that human review of every step is unrealistic once agents are moving money. A signed boolean, carrying the exact condition it was judged against, is one part of that reconstruction, and a firm can produce it without ever having stored the customer's password. It sits in the audit record beside the bank's own logs and the policy decision the bank made; it does not replace them. Every InsumerAPI result is signed twice: ES256 and a post-quantum ML-DSA-65 signature. The time limits above govern whether a bank may act on a result; the signatures stay checkable offline, years later, against a saved copy of the public keys.

We raised the same point with Wells Fargo and Chase's terms in an earlier issue: wanting proof that an agent was revoked is the same ask as wanting proof a delegation is still live. PNC's split between a bank-approved connection and handing over a password has the right shape: authority the customer granted on purpose, kept apart from a secret the agent happens to hold. For developers building the approved-connection side of that split, Skye Meta's access middleware, @skyemeta/access, is a free, MIT-licensed example of the pattern running on InsumerAPI: existing API-key customers pass through unchanged, and an agent can instead authenticate by signing with its wallet. A second door, not a replacement for the first.

Where to Start

If you're building the agent side of this, get a free key at /developers/quickstart/ and build the check in two steps: verify the delegation, then check the proposed transaction against its limits and your own policy before letting it execute. Agents can also skip the key and pay per call through x402. If you're the institution writing this quarter's terms, the clause that assigns blame is the easy part. The harder one is the question a bank should be able to answer before it moves the money: what did the customer authorize this agent to do, and what did we check before we let it act? Read the full argument in the thesis behind this series.

Condition-based access across 37 chains

InsumerAPI: evaluate wallet conditions, get a signed result. No secrets. No identity. Free tier available.

View API Docs

Get every post by email

The Inevitable series and builder notes on condition-based access. Free, one or two posts a week.