Skip to content

armgit / ax

Public Imported from GitHub
0

Agent Executor (AX) is Google's open agentic orchestration runtime

Go

docs: add SECURITY.md with reporting process and current trust boundary (#480)

Maya Wange5fbde792c04
diff --git a/SECURITY.md b/SECURITY.md
new file mode 100644
index 0000000..57103bd
--- /dev/null
+++ b/SECURITY.md
@@ -0,0 +1,17 @@
+# Security Policy
+
+Please use https://g.co/vulnz to report security vulnerabilities.
+
+We use https://g.co/vulnz for our intake and triage. For valid issues we will do coordination and disclosure here on GitHub (including using a GitHub Security Advisory when necessary).
+
+The Google Security Team will process your report within a day, and respond within a week (although it will depend on the severity of your report).
+
+## Current trust boundary
+
+AX is in heavy development. Until authentication and isolation land, treat a deployment as follows:
+
+- The `ax-server` API has no authentication or authorization. Anyone who can reach it can read, create, suspend, resume and delete any Task, Workspace or Model. Only expose it to trusted clients.
+- Atespaces organize resources. They are not a tenant isolation boundary.
+- Setting `spec.debug: true` turns on the guest services, which allow arbitrary command execution and file access inside the sandbox. Whoever can create a Task can set it.
+- Creating a Task injects the Gemini API key (the `gemini-api-secret` secret in the task's atespace, or the server's own `GEMINI_API_KEY`) into the sandbox environment, so anyone who can create a Task can read it.
+- Anything inside a task's sandbox, including the agent and the task runner, should be treated as untrusted. Don't rely on values reported from inside the sandbox for security decisions.