If you discover a security vulnerability in NeuroVerse Governance, please report it responsibly.

Do not open a public GitHub issue for security vulnerabilities.

Instead, email: security@neuroverseos.org

Include:

• Description of the vulnerability
• Steps to reproduce
• Affected versions
• Potential impact

We will acknowledge receipt within 48 hours and provide a detailed response within 7 days.

The following are in scope for security reports:

• Guard engine bypass — any input that should be BLOCKED but isn't
• Prompt injection evasion — patterns that bypass the 63+ safety checks
• World loader path traversal — reading files outside the world directory
• MCP server vulnerabilities — command injection, buffer overflow, unauthorized access
• Playground XSS — cross-site scripting in the web UI
• Plan enforcement bypass — circumventing plan constraints

• Denial of service via extremely large inputs (we enforce size limits, but resource exhaustion on the host machine is out of scope)
• Vulnerabilities in dependencies (report to the dependency maintainer)
• Social engineering

NeuroVerse Governance is built on these security principles:

1. Deterministic evaluation — no LLM in the guard loop. Same input = same verdict.
2. Zero runtime dependencies — no supply chain attack surface.
3. Input validation — all inputs are length-limited before regex evaluation.
4. Fail-closed by default — if the engine can't evaluate, it returns ERROR, not ALLOW.
5. No secrets in code — API keys read from environment variables, config stored with 0600 permissions.