Alexa Trust Intelligence

Alexa+ can spend money by voice, so scammers will target it. This checkpoint stops fake invoices, fake IT calls and pressured purchases before Alexa+ acts.

Comment

Inspiration

Alexa+ can pay vendors, approve invoices and buy with a card on file, by voice, with no screen to double-check. That makes it a new target for old scams. I wanted a checkpoint that runs before Alexa+ acts, not after the money moves.

What it does

A self-hosted MCP server that Alexa+ calls before it pays, approves or shares access. It returns PROCEED, REVIEW or HOLD with reasons. It stops a pressured purchase, a fake invoice with swapped bank details, and a fake IT caller who retries with a new story. Nothing escalates without a real person saying yes.

How we built it

Python MCP server with eight tools. Decisions are fixed, auditable rules. Amazon Bedrock (Nova Lite) reads wording, extracts details from free text and explains results. It never decides. SQLite memory catches repeat attackers. An experimental Gmail button sends an open email through the same checkpoint.

Challenges we ran into

No Alexa+ preview access, so a web app simulates the client. The checks and decisions are real. The SDK's auth sent a header Alexa+ forbids, so I wrote custom middleware.

Accomplishments that we're proud of

No AI model makes the final call. On my own 48 test scenarios: 32/32 correct on money, 15/16 on access. Test results, not production numbers.

What we learned

For money and access, let AI read and explain, and let fixed rules decide.

What's next for Alexa Trust Intelligence

The real Alexa+ connection, shared memory in DynamoDB, and a learned risk signal beside the rules.

Built With

Share this project:

Updates

Submission history