If you discover a security vulnerability in Agent Passport:

1. Do not open a public GitHub issue with exploit details.
2. Instead, email the maintainer directly (or use GitHub Security Advisories if enabled).
3. Please include:
   • A description of the vulnerability
   • Steps to reproduce
   • Any potential impact you see
   • Suggested remediation, if you have one

We will:

• Acknowledge receipt as quickly as possible.
• Investigate and, if confirmed, work on a fix.
• Coordinate a responsible disclosure timeline where appropriate.

Agent Passport is designed to mitigate:

• Malicious Passports: schema/semantic validation and permission policies.
• Malicious tools: virtualized tools for destructive operations in tests.
• Prompt injection attempts: behavior contracts + tool/permission enforcement.
• Sandbox escape attempts: Docker sandbox executor (non-root, resource limits, no host Docker socket).
• Secret exfiltration: redaction in traces, deny-by-default network policies for some runtimes.
• Forged evidence: SHA-256 hashes and Ed25519 signatures on evidence bundles.

• It does not magically make untrusted LLM providers safe.
• It does not prevent model-level jailbreaks; it detects violations when observable via behavior contracts and traces.
• It does not provide full multi-tenant isolation; treat the stack as a component within your broader secured environment.

• Run sandboxed execution workers in restricted environments (e.g. separate node pool, locked-down network).
• Use separate credentials per runtime adapter, scoped with least privilege.
• Always scan container images and dependencies in CI.