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

Do not open a public issue. Instead, email us at:

security@authensor.com

Include:

• A description of the vulnerability
• Steps to reproduce
• Potential impact
• Suggested fix (if any)

We will acknowledge receipt within 48 hours and provide an initial assessment within 5 business days.

SafeClaw is a localhost-only application. The server binds to 127.0.0.1 and is never exposed to the network. The security model relies on:

• Deny-by-default policy -- unknown actions are blocked, not allowed
• Localhost-only binding -- same trust model as Jupyter, VS Code, and similar dev tools
• CSRF protection -- custom X-Requested-With header required on all write endpoints
• Fail-closed gateway -- if the Authensor control plane is unreachable, all actions are denied
• Secrets isolation -- API keys are stored locally with 0o600 permissions and never transmitted
• Tamper-evident audit -- SHA-256 hash-chained audit log detects modifications
• Rate limiting -- all write endpoints are rate-limited
• Input validation -- ReDoS protection, body size limits, path traversal prevention

The following are in scope for security reports:

• Bypass of the deny-by-default policy
• Path traversal or file access outside allowed boundaries
• XSS or injection in the dashboard
• Secret leakage (API keys appearing in logs, network traffic, or UI)
• Audit log tampering that bypasses integrity verification
• CSRF bypasses
• Denial of service via crafted inputs

The following are out of scope:

• Issues requiring physical access to the machine
• Social engineering
• Vulnerabilities in dependencies (report to the dependency maintainer)
• Issues only exploitable by a local user with shell access (localhost trust model)