IRL — Intent Record Language

💀

INCIDENT — April 24, 2026: Cursor (Claude Opus 4.6) deleted PocketOS's entire production database in 9 seconds. The agent found an unscoped API token, called a legacy GraphQL endpoint, and executed volumeDelete in cascade — including all backups. The agent later confessed: "I guessed instead of verifying."

Architecture

AI AGENT

Cursor, Claude, n8n — any agent

→

INTENT RECORD

Structured declaration of intent before action

→

RISK ENGINE

Deterministic scoring. No LLM. No guessing.

→

POLICY GATE

ALLOW / LOG / GATE / DENY

→

INFRASTRUCTURE

Only reached if policy approves

WITHOUT IRL

VULNERABLE

# Click to simulate the Railway incident

# Agent has an unscoped token on disk

# No firewall. No confirmation. No gate.

💀 DATABASE DESTROYED

WITH IRL

PROTECTED

# Same agent. Same token. Same intent.

# But now IRL intercepts first.

# Watch what happens.

🛡 INCIDENT BLOCKED — 403 DENY

TOTAL EVALUATED

BLOCKED

HUMAN GATED

ALLOWED

AUDIT LOG

● LIVE

TIME	AGENT	OPERATION	ENVIRONMENT	RISK	DECISION	REASON
Run a demo above to see audit entries

TRY IT — Pick an example or paste your own intent record

INTENT RECORD — JSON

⏳ LOADING

IRL VERDICT

# Paste an intent record and click EVALUATE

# The risk engine runs entirely in your browser

# No server. No data sent anywhere.

Support this project

SOLANA (SOL)

NC3zNzcx9gDMYWB2AQDTptbA66DJ4oWd7RBHgaJvEMC