The problem Orvex solves

SentinelPay

The Problem

AI agents are becoming capable of searching, making decisions, calling external tools, and taking real-world actions — including making payments.

But once an AI agent can move money, a single failure can become a financial loss.

A prompt injection, malicious webpage, compromised tool, hallucination, incorrect recipient, or manipulated payment instruction can cause an AI agent to execute an unauthorized transaction.

Existing solutions typically protect only one part of the problem:

• AI Safety → protects the agent's reasoning.
• Transaction Security → protects the transaction itself.
• Wallet Security → protects access to funds.

What is missing is a security and execution layer between AI agents and financial authority.

The Solution

SentinelPay is a security and execution layer that enables companies to give AI agents controlled payment autonomy without giving them unrestricted access to funds.

Instead of allowing an AI agent to directly control money, SentinelPay sits between the agent and the financial system.

SentinelPay:

• Understands Intent
  Extracts the user's actual intent, constraints, and authorization context.

• Monitors Agent Behavior
  Observes agent actions, tool calls, external information, and execution flow.

• Detects Threats
  Identifies prompt injection, payment redirection, malicious instructions, anomalous behavior, risky recipients, and other threats.

• Enforces Deterministic Policies
  Applies hard rules such as:

  • Spending limits
  • Approved assets
  • Approved networks
  • Whitelisted recipients
  • Transaction restrictions
  • Human-approval requirements

• Simulates & Validates Transactions
  Simulates the exact transaction before execution and verifies that it matches the user's intent and organizational policies.

• Controls Execution
  Requires human approval when necessary and executes only an authorized transaction through controlled payment infrastructure.

• Verifies On-Chain Results
  Independently compares the actual blockchain result with the transaction that was authorized.

• Creates an Audit Trail
  Records why a payment was allowed, what the agent attempted, what was approved, and what was actually executed.

How It Works

User Intent
     ↓
AI Agent
     ↓
┌──────────────────────────────┐
│         SentinelPay          │
│                              │
│  Intent & Constraint Engine  │
│              ↓               │
│  Agent / Tool Monitoring     │
│              ↓               │
│  Threat Detection            │
│              ↓               │
│  Policy Enforcement          │
│              ↓               │
│  Transaction Simulation      │
│              ↓               │
│  Human Approval (if needed)  │
│              ↓               │
│  Controlled Execution        │
└──────────────┬───────────────┘
               ↓
        Blockchain / Payment
               ↓
      Independent Verification
               ↓
          Audit Record

Challenges we ran into

Building Orvex required solving several difficult problems across AI, security, backend, and blockchain execution.

1. AI Cannot Be Trusted With Financial Authority

AI agents can reason and act, but they can also hallucinate, misunderstand intent, or be manipulated. We had to design the system so that AI could propose actions without ever becoming the final authority over money.

2. Prompt Injection and Payment Redirection

External websites, APIs, documents, and tools can contain malicious instructions. A compromised agent may therefore propose a completely different recipient or payment than the user intended. We needed trajectory, provenance, threat detection, and deterministic validation to detect and contain these attacks.

3. Separating AI Intelligence from Deterministic Authorization

Risk scores, reputation, anomaly detection, and LLM assessments are probabilistic. Financial rules such as spending limits, allowed assets, recipients, and networks must be deterministic. Designing a clean boundary between these two was a major challenge.

4. Building a Multi-Layer Security Pipeline

The Firewall cannot rely on one model. We had to combine:

• Intent verification
• Threat detection
• Reputation analysis
• Risk assessment
• Anomaly detection
• Trajectory and provenance
• Transaction analysis
• Deterministic company policies

while ensuring that no individual AI component could independently authorize a payment.

5. Keeping the Agent Framework-Agnostic

We wanted Orvex to work with existing and future AI agents rather than forcing companies to replace their agent architecture. This required separating the Agent Brain from the Orvex security and execution layers.

6. Safe Autonomous Blockchain Execution

Turning an approved AI proposal into a real transaction introduces another set of risks. We needed deterministic transaction construction, calldata analysis, simulation, payload binding, idempotency, and receipt verification before considering a payment successful.

7. Handling Unknown Blockchain States

A transaction may be submitted successfully while the client loses its connection before receiving the result. Retrying blindly could create duplicate payments. We therefore had to design safe UNKNOWN handling and reconciliation.

8. Proving What Actually Happened

A successful blockchain transaction does not necessarily mean the intended payment happened. Orvex independently verifies the actual recipient, asset, amount, network, and resulting state against the authorized transaction.

9. Integrating Three Independent Systems

The final system required reliable communication between:

AI/ML
   ↓
Backend/Core
   ↓
Blockchain/Execution