openclaw-air-trust

The EU AI Act compliance plugin for OpenClaw — tamper-evident audit trails, consent gating, data tokenization, and prompt injection detection for autonomous AI agents.

Why This Exists

OpenClaw agents can read your email, execute shell commands, send messages, and manage files — autonomously. When something goes wrong, there's no tamper-evident record of what happened, no approval gate for destructive actions, and no protection against prompt injection attacks.

This plugin fixes that. EU AI Act enforcement begins August 2026. This is the compliance layer.

Install

npm install openclaw-air-trust

Then add to your OpenClaw config:

{
  "plugins": ["openclaw-air-trust"]
}

What It Does

| Capability | What It Does | EU AI Act Article | |---|---|---| | Audit Ledger | HMAC-SHA256 tamper-evident chain of every action | Article 12 (Record-Keeping) | | Consent Gate | Blocks destructive tools until user approves | Article 14 (Human Oversight) | | Data Vault | Tokenizes API keys, PII, credentials before they reach the LLM | Article 10 (Data Governance) | | Injection Detector | Scores inbound messages for 15+ prompt injection patterns | Article 15 (Cybersecurity) | | Risk Classifier | Classifies every tool by risk level (CRITICAL/HIGH/MEDIUM/LOW) | Article 9 (Risk Management) | | Compliance Scanner | Checks agent code against all 6 EU AI Act articles | Articles 9-15 |

Plugin Tools

Once installed, these tools are available to your OpenClaw agent:

| Tool | Description | |---|---| | air_audit_status | Get audit chain length, validity, and time range | | air_verify_chain | Verify tamper-evident chain integrity | | air_scan_injection | Scan text for prompt injection patterns | | air_classify_risk | Classify a tool by EU AI Act risk level | | air_export_audit | Export the full audit chain as JSON | | air_compliance_check | Run EU AI Act compliance check on code |

How It Works

Audit Ledger (Article 12)

Every tool call, LLM interaction, consent decision, and injection detection gets appended to a tamper-evident chain:

Entry 1 → hash₁ ──┐
Entry 2 → hash₂ (prevHash = hash₁) ──┐
Entry 3 → hash₃ (prevHash = hash₂) ──┐

Each entry is signed with HMAC-SHA256. Modifying any record breaks the entire chain downstream.

Consent Gate (Article 14)

When the agent tries to call a destructive tool, the consent gate intercepts and sends an approval request. Risk classification is built-in: critical (code execution), high (file writes, deploys), medium (network/email), low (reads).

Data Vault (Article 10)

Before tool arguments or context reaches the LLM, the vault scans for sensitive patterns and replaces them with opaque tokens. 14 built-in patterns: OpenAI/Anthropic/AWS/GitHub/Stripe keys, emails, phone numbers, SSNs, credit cards, connection strings, bearer tokens, private keys, and password assignments.

Injection Detector (Article 15)

Scans inbound messages for 15+ prompt injection patterns: role override, identity hijacking, privilege escalation, safety bypass, jailbreak, data exfiltration, encoding evasion, and more. Three sensitivity levels (low/medium/high) control which patterns are active.

Configuration

{
  "plugins": {
    "openclaw-air-trust": {
      "enabled": true,
      "gatewayUrl": "https://your-air-gateway.example.com",
      "gatewayKey": "your-api-key",
      "consentGateEnabled": true,
      "consentAlwaysRequire": "exec,spawn,shell,deploy",
      "consentRiskThreshold": "high",
      "consentTimeoutMs": 30000,
      "injectionEnabled": true,
      "injectionSensitivity": "medium",
      "injectionBlockThreshold": 0.8,
      "vaultEnabled": true
    }
  }
}

Standalone Usage

You can also use the components directly without OpenClaw:

import { createAirTrustPlugin } from 'openclaw-air-trust/standalone';

const trust = createAirTrustPlugin({
  enabled: true,
  consentGate: { enabled: true, alwaysRequire: ['exec', 'deploy'] },
  injectionDetection: { enabled: true, sensitivity: 'medium', blockThreshold: 0.8 },
});

Part of the AIR Blackbox Ecosystem

| Package | What It Does | |---|---| | air-blackbox-mcp | MCP server for Claude Desktop — 10 compliance tools | | air-langchain-trust | Python trust layer for LangChain agents | | air-crewai-trust | Python trust layer for CrewAI agents | | air-autogen-trust | Python trust layer for AutoGen agents | | openclaw-air-trust | ← You are here |

Learn more at airblackbox.ai

Development

git clone https://github.com/airblackbox/openclaw-air-trust.git
cd openclaw-air-trust
npm install
npm run build
npm test

License

Apache 2.0