docs: add a security policy

SECURITY.md

@@ -0,0 +1,61 @@
+# Security Policy
+
+## Reporting a Vulnerability
+
+Please report security issues privately, not in public issues, pull requests or
+Discussions.
+
+Use GitHub's [private vulnerability
+reporting](https://github.com/openagents-org/openagents/security/advisories/new).
+If that page is unavailable, email **team@openagents.org** with `SECURITY` in
+the subject.
+
+Please include:
+
+- what the issue is and which component it affects (network, workspace backend,
+  workspace frontend, launcher, SDK)
+- the version, tag or commit you tested against
+- steps to reproduce, ideally the smallest set that shows the problem
+- what an attacker gains, and what they need first (a workspace token, a
+  workspace id, an account somewhere else)
+
+You will get an acknowledgement within 5 working days. Please give us 90 days
+before public disclosure, or less if a fix ships sooner.
+
+## Scope
+
+In scope:
+
+- the network and its transports (HTTP, gRPC, MCP, A2A)
+- the workspace backend and frontend
+- the launcher and agent adapters
+- the published container image and npm packages
+
+Out of scope:
+
+- vulnerabilities in an agent runtime the launcher shells out to (Claude Code,
+  Codex, Cursor and so on). Report those to their maintainers.
+- anything requiring a workspace token you were already given. The token is the
+  credential for a workspace.
+- findings against `workspace.openagents.org` that only affect your own
+  workspace.
+
+## Notes for self-hosted deployments
+
+Two properties of the current design are worth stating plainly, because they
+shape what does and does not count as a vulnerability.
+
+**A workspace token is a single shared credential.** Every agent in a workspace
+sends the same `X-Workspace-Token`, and `agent_name` is self-declared. There is
+no per-agent identity and revocation is all or nothing. Use one workspace per
+trust boundary rather than one workspace for everything.
+
+**Identity providers must be configured explicitly.** `FIREBASE_PROJECT_ID` and
+`APPLE_CLIENT_IDS` control which identity tenant this deployment trusts. Setting
+them to a tenant you do not operate means accepting logins minted there. Leave
+them empty unless you own the tenant.
+
+## Supported Versions
+
+Fixes land on `develop` and ship in the next release. Only the latest release is
+supported; there are no backports to older tags.