Limiting AI agent access to bank accounts over MCP
On this page
Unvetted AI agents reach a bank account only when someone hands them a working credential. This page cites no public report on any specific incident, so it makes no claim about how many accounts were affected. The practical goal for developers is to stop an agent from holding more bank access than its task needs, and scoped tokens, explicit consent, and an audit trail cover that goal.
How an AI agent gets access to a bank account
An AI agent touches a bank account through one of three credential paths. Each path decides how much control the account owner keeps.
- An OAuth grant that the account owner approves on a consent screen.
- A token issued by an account aggregator that the developer's app holds on the owner's behalf.
- A stored API key or password that a script reads from a configuration file.
The third path gives the least control. The credential keeps working after the owner forgets the agent exists, because nothing expires it. The first path gives the most control: the owner sees the requested scopes and can revoke them from the bank's own settings.
Check the agent and its tools before granting anything
Start with the tool, not the model. A tool that reads balances needs a different grant than a tool that starts transfers. Write down each action the agent may take before you connect it to any account.
For every MCP tool in the chain, record four facts:
- The publisher of the tool and the contact for security reports.
- The scopes the tool requests from the bank or aggregator.
- The servers that receive account data, and where those servers run.
- Whether the publisher gives you access to its request logs.
A tool that cannot answer these four questions does not belong in a path that touches money. Treat a missing answer as a reason to pause, not as a detail to fill in later.
Scope credentials to read-only and to one account
Read-only access covers most agent tasks. An agent can summarise spending, flag a missed bill, or reconcile categories without any write permission. Grant write scopes only when a named task requires them.
Bind each token to a single account. A token that spans every account the owner holds turns one mistake into a wide exposure. Set the expiry to match the task length, not to a year.
- Read balances and transactions for one account.
- Reject transfer and payee-management scopes unless a named task needs them.
- Set an expiry that ends when the task ends.
Review each scope once a month. Revoke any scope that no task uses any longer.
Require a person to approve money movement
Put a confirmation step between the agent and any payment. The agent proposes a transfer with the amount, the payee, and the reason. A person approves the proposal in a separate interface. The agent never receives the approval credential.
This pattern adds a few seconds to each payment. It also stops a prompt injection from turning a read task into a payment. An agent that reads an invoice full of hidden instructions can propose a transfer, but it cannot complete one without approval.
Keep the approval interface outside the agent's runtime. If the agent can write to the approval queue, the approval step protects nothing.
Common mistakes that widen access
Most over-broad access comes from convenience, not from intent. A developer reuses a personal token for a demo. A broad scope gets approved once during testing and never reviewed again. A shared secret ends up in a repository that many people can read.
- Reusing a personal token in an agent's configuration.
- Keeping a broad test scope in production.
- Storing a credential in a file the agent can read.
Each mistake turns a narrow task into a wide grant. A fixed review schedule catches most of them before they cause harm.
Log every call and keep a kill switch
Record the tool name, the scope used, the account touched, and the result for each call. Keep these logs in storage that the agent cannot write to. An agent that can edit its own logs can also hide its own actions.
A kill switch is one action that revokes every token for one agent. Test it before you need it. A kill switch that nobody has run tends to fail at the moment it matters.
Set an alert for any call outside the agent's normal pattern. A new account, a new scope, or a call at an unusual hour deserves a look before the next call runs.
Where MCP tools fit in this setup
An MCP gateway sits between the agent and the tools it calls. The gateway can hold credentials on the server side, so the model never sees raw secrets. The model sends a tool request, and the gateway attaches the authorised token for that single call.
APIbase lists MCP tools across a large and growing catalog, with hundreds of providers and tools across many categories. Check the catalog for a specific banking or finance provider before you assume one exists. Each listing shows its published price, which ranges from 0 to 29.99 USD.
A gateway does not replace the controls above. Scoped tokens, approval steps, and logs still belong in your own design. The gateway only keeps the secrets out of the model's reach.
What to do if an agent already has bank access
Start by listing every active token tied to the agent. Include tokens held by your app, by any aggregator, and by any script in your repository.
Revoke any token with write scope that the current task does not need. Then review the logs for transfers or payee changes that nobody requested. If a log shows an unexpected call, revoke all tokens for that agent, then rotate the credentials it could read.
Notify the account owner before you resume access. The owner decides whether the agent keeps any access at all.
Live pricing — developer
Data source: live tool catalog snapshot, fetched 2026-11-16.
Chart data as a table
| Tool | Price per call | Cache-hit price |
|---|---|---|
| Usage Time Series | $0.0 | $0.0 |
| Per-Tool Usage Breakdown | $0.0 | $0.0 |
| Usage Summary | $0.0 | $0.0 |
| Discover Tools | $0.0 | $0.0 |
| Batch Tool Calls | $0.0 | $0.0 |
| Tool Quality Metrics | $0.0 | $0.0 |
| Tool Quality Rankings | $0.0 | $0.0 |
| List Programming Languages | $0.001 | $0.0001 |
| Check CVE ID Reservation Status | $0.001 | $0.0001 |
| Security Advisories (deps.dev) | $0.001 | $0.0001 |
| Dependency Tree (deps.dev) | $0.001 | $0.0001 |
| Package Info (deps.dev) | $0.001 | $0.0001 |
| Tool | Provider | Price/call | Cache-hit |
|---|---|---|---|
| Usage Time Series | account | $0 | $0 |
| Per-Tool Usage Breakdown | account | $0 | $0 |
| Usage Summary | account | $0 | $0 |
| Discover Tools | apibase | $0 | $0 |
| Batch Tool Calls | platform | $0 | $0 |
| Tool Quality Metrics | platform | $0 | $0 |
| Tool Quality Rankings | platform | $0 | $0 |
| List Programming Languages | judge0 | $0.001 | $0.0001 |
| Check CVE ID Reservation Status | cve-mitre | $0.001 | $0.0001 |
| Security Advisories (deps.dev) | depsdev | $0.001 | $0.0001 |
| Dependency Tree (deps.dev) | depsdev | $0.001 | $0.0001 |
| Package Info (deps.dev) | depsdev | $0.001 | $0.0001 |
Connect via MCP
Data source: gateway design, no measured values.
$ curl -X POST https://apibase.pro/api/v1/tools/account.timeseries/call \
-H "Content-Type: application/json" -d '{"params": {}}'FAQ
Does this page confirm that unvetted AI agents are accessing US bank accounts?
No. This page cites no public report on the incident and verifies no source for it. Developers can still limit their own exposure by scoping tokens to read-only access, requiring approval for payments, and keeping logs that the agent cannot edit.
Can an AI agent move money without a person approving it?
Yes, if the agent holds a token with transfer scope. Remove that scope, or route every transfer through a separate approval step that the agent cannot reach.
Is read-only access enough for most agent tasks?
For many tasks, yes. A read-only token lets an agent summarise balances and transactions. It blocks payments and changes to account settings.
How long should a bank token last for an agent?
Match the expiry to the task. A token for one reconciliation run should expire when that run ends, not weeks later.
Recommended next step
- STEP 1/2 Connect your MCP client
- STEP 2/2 Browse the full catalog
Related guides
Written with AI assistance from APIbase's published documentation and public data; checked automatically against them. Corrections: https://apibase.pro/
Published: 2026-11-16Source: live apibase.pro catalog snapshot