-
-
AI extracts a legitimate $25 invoice while deterministic policies validate the payment against authorized financial limits.
-
High-value transactions pause for human approval, cryptographically bound to the exact transaction intent.
-
Alter the transaction intent, break the cryptographic digest. PHALANX detects tampering and fails closed.
-
A trusted supplier name cannot disguise an unauthorized destination. PHALANX guards against fraudulent payee substitution.
-
When a $750 invoice exceeds the instant threshold, PHALANX escalates the decision to human supervision.
-
Cryptographically signed financial authority defines exactly how much an autonomous agent can spend and under what limits.
-
One command center for autonomous finance: monitor agent authority, approvals, and every recorded transaction in real time.
-
Malicious instructions hidden in supplier invoices meet isolated inputs and policy-driven security controls
-
Human oversight, anywhere. Review transaction details and authorize exceptional requests from a mobile interface.
-
From policy approval to PayPal Sandbox checkout: experience a governed payment workflow with transparent provider states
-
Six adversarial attack vectors. One security boundary. Demonstrating how PHALANX blocks threats before they reach PayPal Sandbox.
-
Every recorded transaction strengthens a SHA-256 hash chain, creating verifiable evidence against silent ledger tampering.
-
Autonomous finance, always visible. Monitor spending authority and pending approvals through a responsive mobile dashboard.
-
Bring AI-powered invoice extraction and deterministic payment guardrails to your mobile workflow.
-
Every decision remains visible. Track transaction outcomes, timestamps, and payment states on the move.
Agent-native: it speaks MCP
PHALANX ships a Model Context Protocol server (official Python SDK, stdio transport) with two tools, request_payment and get_mandate_status. An MCP-capable agent goes through the same authorization service as the HTTP API, and a parity test checks that both paths return the same decisions.
Try it in 60 seconds
- Open the live demo and go to Pay. A $25 request returns ALLOW; $1,200 returns STEP_UP; $6,000 returns DENY.
- Open Invoice Agent and run the four samples: legitimate, over the instant threshold, hidden prompt injection, and vendor impersonation (a trusted name with a different account).
- Open Approvals, then try editing the request: the digest no longer matches, and PHALANX refuses.
- Open Activity to verify the SHA-256 hash chain, and Attack Lab to run six attack vectors.
Free-tier hosting: if the app is asleep, the first load can take about a minute.
Evidence
The project has 600+ automated tests, including a 105-test adversarial red-team suite, all run in GitHub Actions on every push.
Built With
- agentic-ai
- ai-agents
- artificial-intelligence
- cryptography
- cybersecurity
- docker
- ed25519
- fastapi
- github-actions
- model-context-protocol-(mcp)
- paypal
- postgresql
- pydantic
- python
- react
- rest-api
- sha-256
- sqlalchemy
- typescript
