Skip to content

Security: ml8s/agentgo

Security

SECURITY.md

Security Policy

agentgo is an internal runtime and must not be exposed directly to the public internet.

Required controls:

  • Internal bearer token between the gateway and agentgo outside local development.
  • A dedicated internal ops listener for /metrics; pprof is disabled unless explicitly enabled for debugging. Network isolation and optional gateway-owned authentication are deployment responsibilities.
  • Verified user context supplied by the gateway for AG-UI.
  • The browser identity header is adopted on AG-UI only; A2A machine callers sharing the internal token cannot spoof a user identity.
  • Direct mutual-TLS callers are identified by their verified certificate subject by default. Trusting relayed identity headers on that listener (AGENTGO_MTLS_IDENTITY_SOURCE=header) is an explicit deployment opt-in for gateway peers and must be documented as such.
  • Internet-facing request rate limiting is owned by the gateway; the runtime's internal per-source limiter only resists token brute force.
  • Bounded per-source throttling of invalid bearer or identity attempts.
  • Deployment integrity is owned by deployment system, which assembles the images by commit sha; the runtime keeps no duplicate artifact pin.
  • HTTPS-only LLM endpoints outside development and no credentials embedded in configured URLs.
  • Explicit per-Agent MCP tool allowlists.
  • No secrets, prompts, raw model output, or tool payloads in logs.
  • Least-privilege, non-root deployment.

Reporting a vulnerability

Do not open a public issue for a security vulnerability. Use GitHub's private vulnerability reporting for this repository.

There aren't any published security advisories