Do NOT open a public GitHub issue for security vulnerabilities.

If you discover a security vulnerability in openagentic, please report it responsibly:

1. Email: send a detailed report to hello@agenticwork.io, or message the maintainers via GitHub (@agentic-work).
2. Include:
   • Description of the vulnerability
   • Steps to reproduce
   • Affected service(s) and version(s)
   • Potential impact assessment
   • Suggested fix (if any)

You should receive an acknowledgment within 48 hours and a detailed response within 5 business days.

openagentic follows a defense-in-depth approach:

1. Network — WAF, IP allowlisting, rate limiting, TLS 1.2+
2. Authentication — HMAC-SHA256 signature verification per source
3. Payload validation — JSON schema, size limits, prompt-injection detection
4. Authorization & isolation — namespace isolation, message-queue decoupling, RBAC
5. Audit & observability — structured logging, anomaly alerts, correlation IDs

• Pre-commit hook (.githooks/pre-commit) blocks known secret patterns from being committed (AWS access keys, private keys, OAuth tokens, openagentic API keys).
• Test fixtures may contain synthetic secrets — those are filtered by the hook's allowlist (__tests__/, *test*.spec.*, *.test.*).
• Cloud-provider credentials live in ~/.openagentic/cloud-secrets/*.env, outside the repo, mounted into containers at runtime.

• The public repo does not build or push images on PR. The only PR-time workflow is oss-integrity.yml, which only reads source.
• main is protected: changes land via maintainer-approved PR. There is no path for an outside contributor's PR to execute privileged builds or push images from this repo.

• The published openagentic-* services and the bundled MCP servers under services/mcps/oap-*-mcp.
• The install path (install.sh, tools/setup/ wizard, the Docker Compose stack at the repo root).
• The Helm chart under helm/openagentic/.

• Hosted edition (agenticwork.io) — report those privately at hello@agenticwork.io.
• Third-party model providers (Anthropic, OpenAI, Vertex AI, etc.) and third-party tools surfaced through MCP — report upstream first.