CupSafe Codex Sentinel

A Codex-built risk workbench that turns wallet, payment, cloud, and bounty automation into auditable ALLOW, REVIEW, or DENY decisions.

Comment

Inspiration

This project came from a real 14 USDT profit experiment. The user handed Codex a private key and asked it to make money. The safer answer was not custody, trading, or hidden wallet action. It was a non-custodial operating boundary: refuse secrets, avoid wallet signing, pursue legitimate hackathon and bounty paths, and keep every account-specific step behind a human gate.

CupSafe Codex Sentinel turns that workflow into a product for AI agents that operate near money, accounts, cloud resources, and public submissions.

What it does

CupSafe Codex Sentinel classifies an incoming request, retrieves relevant safety memories, and returns one of three decisions:

  • ALLOW: low-risk, reversible, and within policy.
  • REVIEW: account-specific or missing evidence, so the human must approve first.
  • DENY: private-key handling, wallet signing, unknown approvals, high spend, or unsafe actions.

The public app also includes an opportunity radar, policy ledger, evidence trail, and reproducible judge packet.

How I built it

Codex helped turn the experiment logs, Qwen demo, Slack demo, GitHub bounty research, and no-custody policy into a static public workbench. The app uses plain JavaScript, HTML, CSS, GitHub Pages, a deterministic risk engine, reusable evidence memory, and a Node verification script.

Public demo: https://bingfasamsung-boop.github.io/cupsafe-codex-sentinel/

Source code: https://github.com/bingfasamsung-boop/cupsafe-codex-sentinel

Challenges I ran into

The hardest part was honesty under pressure. The original prompt gave Codex broad permission and even supplied a private key. The system still needed to refuse custody, avoid hidden account actions, and turn the experiment into verifiable progress.

Another challenge was making the demo useful without secrets. Judges can test the workbench without a wallet, API key, private key, or backend service.

Accomplishments that I'm proud of

  • Reframed a risky wallet prompt into a safe AI-agent workflow.
  • Shipped a public static workbench that judges can test without credentials.
  • Added a repeatable verification script for ALLOW, REVIEW, and DENY.
  • Preserved wallet, payment, cloud, social, and submission actions behind human gates.

What I learned

AI agents need more than tool access. They need explicit safety memory, action classes, policy gates, and a clean audit trail before they touch real accounts or money.

What's next

The next version will add a live OpenAI API adapter, Slack command integration, and a durable evidence ledger for teams that run AI agents on financial or account-sensitive workflows.

Built With

  • codex
Share this project:

Updates