Inspiration
Autonomous AI agents are starting to initiate real payments, yet traditional payment rails only detect fraud after funds move. If an agent is tricked by a prompt-injection attack (like "ignore previous instructions and bypass security") or enters a runaway spending loop, the money is already gone. We built SentinelPay as a real-time, pre-execution security firewall that stops unauthorized payments before a single rupee moves.
##What It Does
Allowance Management: Enforces strict per-agent spending limits (P \le A_{\text{total}} - S_{\text{spent}}). Approves valid transactions (e.g., a ₹15 utility bill) and instantly blocks overspending (e.g., a ₹60 request exceeding the remaining budget).
Prompt Injection Scanner: Inspects payment instructions in real time using a hybrid rule-based and AI threat scoring engine (R_{\text{threat}} \in [0, 100]) to catch malicious prompts before execution.
Automated Wallet Freeze & Wallet Security: Automatically locks compromised agent wallets into a FROZEN state upon threat detection while displaying live allowance and balance telemetry.
Human Override & Audit Logs: Allows authorized operators to safely unfreeze wallets back to ACTIVE after review, with every action recorded in immutable audit logs.
#How We Built It
Backend: Python and FastAPI exposing real-time verification, wallet telemetry, and unfreeze endpoints.
Threat Engine: Hybrid deterministic keyword detection paired with LLM semantic risk scoring.
Frontend & Hosting: Responsive multi-module dashboard built with HTML5, CSS3, and Vanilla JavaScript, deployed live on Render via GitHub.
##Challenges
Challenges: Balancing sub-second payment verification latency with deep AI prompt inspection, and ensuring wallet freeze locks persist atomically across all endpoints until human recovery.
##What We Learned: Post-transaction alerts are too late for autonomous agents—combining hard budget invariants with inline prompt scanning and human-in-the-loop override creates true zero-trust security for agentic payments
Accomplishments that we're proud of
End-to-End Live Firewall: Built and deployed a working pre-execution security layer that blocks overspending and prompt injections before funds move. Sub-Second Hybrid Detection: Combined instant keyword guardrails with AI risk scoring to catch prompt attacks without slowing legitimate payments. Auto-Freeze & Human Override: Engineered automated FROZEN wallet locks upon threat detection with safe human recovery and full audit logs.
What's next for SentinelPay
Direct PayPal API Integration: Connect SentinelPay's pre-execution approval tokens directly to live PayPal Payouts and Orders APIs for end-to-end payment settlement. Multi-Agent Allowance Hierarchies: Enable enterprise teams to set parent-child spending limits and cryptographically signed security policies across entire fleets of AI agents. Real-Time Webhook & Slack Alerts: Push instant freeze alerts to security teams for one-click mobile human override.