AI and MCP

Is it safe to connect an AI assistant to your finances? A developer's view

A developer's threat model for connecting ChatGPT or Claude to financial data through MCP: scopes, read-only starts, prompt injection and disconnecting.

Editorial illustration for Is it safe to connect an AI assistant to your finances? A developer's view

It can be reasonably safe if you treat it like giving a script an API key: grant the smallest scope that answers your question, start read-only, assume anything the assistant reads can leave your machine, and know how to revoke access. It is not safe to hand over write access on day one.

If you build personal tools, you already think this way about tokens. The same instincts apply to a connector, with a few twists specific to language models.

What is the real threat model?

  • Over-broad scope: the assistant can read or change more than the task needs.
  • Data in the prompt: anything it reads is sent to the assistant provider as part of the conversation.
  • Prompt injection: text inside your data, such as a merchant name or a note, could try to steer the model.
  • Write mistakes: a confident but wrong bulk recategorisation.
  • Forgotten access: a connection you made once and never revisited.

How should I choose a scope?

SpendixAI lets you connect Claude or ChatGPT via MCP and choose a scope, and you can start read-only and disconnect any time. Read-only is the sensible default: it covers questions like "what did I spend on subscriptions in October" without any ability to change a record.

TaskScope neededRisk if it goes wrong
Ask for category totalsread-onlylow: data is read, nothing changes
Draft a budget from historyread-onlylow
Recategorise 40 rowswritemedium: wrong changes, though deleted items sit in a recycle bin for 30 days
Delete or bulk editwritehigher: verify every action first

What does prompt injection look like in a ledger?

Imagine a merchant string that reads like an instruction. A model that treats data as commands might follow it. You cannot fully prevent this, which is another reason to keep scope read-only and to read the assistant's proposed actions before approving anything. Never paste keys or tokens into the chat to fix a connection problem.

How do I test the connection safely?

  1. Create a read-only connection and ask one narrow question.
  2. Compare the answer to a report you trust in the app.
  3. Check the connections list and note what each one can do.
  4. Disconnect, and confirm the assistant can no longer see the data.
  5. Only then decide whether write scope is worth it.

What should I log and review?

Treat the connection like infrastructure you own. Keep a short note of the date you connected, the scope, the client or assistant name, and what you meant to do with it. A month later, the connections list in the app should match your note exactly; any mismatch is the thing to investigate first.

  • Date connected and scope chosen.
  • The question or task it was for.
  • Whether anything was written, and the rows touched.
  • The date you disconnected.

If you ever script around the assistant, keep its output out of anything that auto-applies changes. A human step between the model's suggestion and a write is the cheapest control you will ever add, and it also protects you from the prompt-injection case above.

What does the app not protect against?

The app controls what the connector may touch; it cannot control what you ask or what the assistant provider does with the conversation. SpendixAI says it does not sell personal information and lets owners export data and delete transactions, but those assurances are about the app. Read the assistant provider's own data terms separately, and keep sensitive statements out of chats you do not need.

Finally, record when you connected and what you asked, in the same way you would log a deploy. It makes the review in a month's time quick.

SPENDIXAI JOURNAL

Useful ideas start with a clearer picture of your own spending.

Explore your reports