SentinelX — Zero-Trust Security for AI Payments

Inspiration

The idea for SentinelX started with a simple question we kept coming back to: what happens when AI agents are no longer just recommending what to buy, but actually buying it for us?

While exploring the future of AI-powered commerce, Pranit, Sanvi, and Onkar realized that giving an AI agent access to payments creates a security problem that traditional fraud systems aren't really designed for. An agent could be properly authenticated and still make a dangerous decision because of a prompt injection, a compromised tool, manipulated merchant information, or simply a gradual drift away from the user's original intent.

That led us to a slightly uncomfortable thought: authentication tells us who the agent is, but it doesn't tell us whether we should trust what it is doing right now.

We built SentinelX around that gap.

Instead of treating an AI agent as trusted after a one-time authentication check, SentinelX continuously evaluates its identity, permissions, intent, behavior, instruction provenance, merchant context, and transaction risk before allowing money to move. The goal was not to build another fraud dashboard, but a security layer specifically designed for the new reality of agentic payments.

What We Built

We designed SentinelX as a zero-trust security gateway between an AI payment agent and PayPal. Every payment request passes through deterministic authorization and security analysis before it can reach the payment layer.

We also wanted the project to be more than a collection of security rules. So we built an interactive Attack Lab where we could simulate realistic attacks such as prompt injection, capability escalation, merchant manipulation, replay attacks, and behavioral hijacking, then observe exactly why SentinelX allowed, restricted, or blocked the transaction.

What We Learned

Building SentinelX taught us that securing AI systems requires a different mindset from securing traditional applications. The question is no longer simply "Is this request authenticated?" but also "Why is the agent doing this, where did the instruction come from, is it within the user's intent, and does its behavior still make sense?"

The biggest challenge was balancing AI flexibility with deterministic security. We deliberately kept critical authorization decisions outside the LLM. AI can analyze and explain evidence, but it cannot override a policy, increase its own spending limit, or authorize a payment.

That principle became the foundation of SentinelX: