Skip to main content
Shadow agent quarantine lets you isolate unknown or suspicious agents from making LLM requests. Quarantine is a manual admin action - the same Redis-backed enforcement as the kill switch, but tracked and labeled separately for governance of unregistered or rogue agents.

How It Works

  1. An admin calls POST /agents/{id}/quarantine with a reason after identifying an unknown or suspicious agent.
  2. The API marks the agent quarantined in the database and also writes a block to Redis for instant proxy enforcement.
  3. The proxy blocks all subsequent requests from the quarantined agent with 403 Forbidden.
  4. An admin reviews the agent and calls POST /agents/{id}/release-quarantine to restore access.
Quarantine is not applied automatically today - there is no auto-quarantine of newly discovered shadow agents. An admin decides to quarantine an agent explicitly.

Quarantine an Agent

Release from Quarantine

After reviewing a quarantined agent, release it to restore proxy access:

List Quarantined Agents

What the Agent Sees

When a quarantined agent sends a request through the proxy:

Audit Trail

Quarantine and release actions do not currently emit their own audit events. If you need a record of quarantine actions, track them through your own tooling around the API calls, or query the agent’s current quarantine_reason via GET /agents/quarantined.

Enforcement Architecture

Quarantine uses the same Redis-backed enforcement as the kill switch:
  • Write path: API writes quarantine status to PostgreSQL and publishes a block to Redis.
  • Read path: Proxy checks Redis on every request - quarantined agents get an immediate 403.
  • Fail-open: If Redis is unavailable, the proxy falls back to allowing the request.
  • Latency: Quarantine takes effect within milliseconds of the API call.

Quarantine vs Kill Switch

Use Cases

  • Shadow agent response - Quarantine agents you discover are not in the registry.
  • Onboarding gate - Quarantine new agents until reviewed by security.
  • Incident containment - Quarantine suspicious agents while investigating.
  • Zero-trust enforcement - No agent gets proxy access without explicit registration and review.