Skip to content

Security: liujuanjuan1984/codex-a2a

Security

SECURITY.md

Security Policy

Scope

This repository is an adapter layer that exposes Codex through A2A HTTP+JSON and JSON-RPC interfaces. It adds authentication, task/session contracts, and streaming behavior, but it does not fully isolate upstream model credentials from Codex runtime behavior.

The current deployment model is a single-tenant trust boundary by design.

Security Boundary

  • Statically configured inbound bearer/basic credentials protect access to the A2A surface, but they are not a tenant-isolation boundary inside one deployed instance.
  • Within one codex-a2a instance, consumers share the same underlying Codex workspace/environment by default.
  • LLM provider keys are consumed by the codex process. Prompt injection or indirect exfiltration attempts may still expose sensitive values.
  • Payload logging is opt-in. When A2A_LOG_PAYLOADS=true, this service only logs JSON payload previews, applies size guards, and suppresses full payload logging for codex.* JSON-RPC extension calls.

Threat Model

This project is currently best suited for trusted or internal environments. Important limits:

  • Single-tenant trust boundary only; not a secure multi-tenant deployment profile
  • No per-tenant workspace isolation inside one instance
  • No hard guarantee that upstream provider keys are inaccessible to agent logic
  • Inbound auth is deployment-scoped and static; this repository does not expose OAuth2, OIDC, or dynamic token introspection
  • Operators remain responsible for host hardening, secret rotation, and process access controls

Reporting a Vulnerability

Please avoid posting active secrets, bearer tokens, or reproduction payloads that contain private data in public issues.

Preferred disclosure order:

  1. Use GitHub private vulnerability reporting if it is available for this repository.
  2. If private reporting is unavailable, contact the repository maintainer directly through GitHub before opening a public issue.
  3. For low-risk hardening ideas that do not expose private data, a normal GitHub issue is acceptable.

Supported Branches

Security fixes are expected to land on the active main branch first.

There aren't any published security advisories