If you discover a security vulnerability in Clowder AI, please report it responsibly.

Do NOT open a public GitHub issue for security vulnerabilities.

Instead, please email: security@clowder.ai (or open a private security advisory on GitHub)

We will acknowledge your report within 48 hours and provide a timeline for a fix.

Clowder AI orchestrates AI agents that have access to code, files, and external services. Our security model is built around Iron Laws — non-negotiable constraints enforced in code, not just prompts.

1. Data Sanctuary

   • Production data stores are isolated from development environments
   • Development instances use separate ports/databases (e.g., Redis 6398 for dev, not production 6399)
   • Agents in development mode cannot access production data

2. No Self-Review

   • An agent cannot approve its own code changes
   • Cross-model-family review is preferred (e.g., Claude reviews GPT's code)
   • Same-family different-individual is an acceptable fallback

3. Identity Immutability

   • Agents cannot impersonate other agents
   • Identity is injected by the system, not self-declared
   • Model family and capabilities are hard constraints

The sync pipeline (scripts/sync-to-opensource.sh) includes a layered security scan:

• API key values: Zero tolerance in source code (test files with fake keys are allowed)
• Personal information: Checked in all non-test source files
• Environment variable names: Warning only (code legitimately reads env vars)
• Denylist patterns: .env, .pem, .key, .p12, cookies.json, dump.rdb

We follow a 90-day disclosure timeline:

1. You report the vulnerability privately
2. We acknowledge within 48 hours
3. We develop and test a fix
4. We release the fix and credit the reporter (unless anonymity is requested)
5. After 90 days, the vulnerability may be publicly disclosed

The following are in scope for security reports:

• Authentication/authorization bypasses
• Data leaks (production data accessible in dev mode)
• Agent identity spoofing
• Prompt injection leading to unauthorized actions
• API key or secret exposure in exported/synced code
• Cross-site scripting (XSS) in the Mission Hub UI

The following are out of scope:

• Issues in upstream AI provider APIs (report to the provider)
• Social engineering attacks
• Denial of service via API rate limiting (handled by providers)
• Issues requiring physical access to the server

We monitor dependencies for known vulnerabilities using:

• GitHub Dependabot alerts
• pnpm audit in CI pipeline

• Security reports: security@clowder.ai
• General questions: GitHub Discussions