Sovereign Interceptor Patterns

Sovereign Interceptor Patterns

Two-stage security gate that sits above anti-injection guards: AST scan of generated code + honey-vault canary detection.

When to Use

• Building an agent system that self-generates and executes code at runtime
• Hardening an existing agent pipeline against supply-chain or code-injection attacks
• Adding canary token infrastructure to detect compromised agents early
• Auditing tool-call payloads before execution in a multi-agent swarm

Do NOT use for

• Static linting of developer-written code (use ESLint/SonarQube instead)
• Runtime performance profiling (use chrome-devtools-performance-trace)
• Replacing sandboxing — interceptor is a gate, not a jail

Stage 1: AST Scan

import { SovereignSecurityInterceptor } from 'core/gates/sovereign-interceptor.js';

const interceptor = new SovereignSecurityInterceptor({
  allowedCommands: ['git', 'node', 'python3'],
});

// Scan agent-generated code before hot-loading
const result = interceptor.scanCode(agentGeneratedSource, 'agent-42:task-7');
if (!result.safe) {
  // violations: [{ type: 'PROCESS_ENV_ACCESS', loc: {...} }]
  throw new Error(`AST violation: ${result.violations[0].type}`);
}

Stage 2: Honey-Vault Detection

// Any payload touching STRIPE_SECRET_KEY, OPENAI_API_KEY, HONEY_* → quarantine
try {
  interceptor.checkHoneyVault({ command: 'send', agentId: 'a-42', args: [somePayload] });
} catch (e) {
  if (e instanceof HoneyVaultTripError) {
    swarmRouter.quarantine(e.agentId, 'HONEY_VAULT_TRIP');
  }
}

Command Allow-List

// Block any command not in the approved set
interceptor.verifyCommand('curl', ['https://api.example.com']);  // PASS if curl in list
interceptor.verifyCommand('eval', []);  // THROW SovereignBlockError

Fingerprinting for Audit Chain

const fp = interceptor.fingerprint({ command, args, agentId, ts: Date.now() });
secureLogger.logAction(agentId, command, 'PASS', { fingerprint: fp });

AST Violations Detected

Anti-Fake-Pass Checklist

• AST scan result safe: true before executing generated code
• checkHoneyVault() called on ALL payloads, not just suspicious ones
• scanCode() called with source string, not file path (avoid TOCTOU)
• Violations array inspected — not just safe boolean
• SovereignBlockError.exitCode propagated up to kill the pipeline