fix: Security vulnerability of sudo session enablement (#2162)

<!--
Please precisely, concisely, and concretely describe what this PR changes, the rationale behind codes,
and how it affects the users and other developers.
-->

Fixes #2151, follow-up of #1530.

This PR fixes the security vulnerability in  that was implemented using environment variables.

**Checklist:** (if applicable)

- [x] Milestone metadata specifying the target backport version
- [x] Mention to the original issue
- [x] Installer updates including:
  - New mandatory config options

Backported-from: main (24.09)
Backported-to: 24.03
Backport-of: 2162

changes/2162.fix.md

@@ -0,0 +1 @@
+Fix security vulnerability for `sudo_session_enabled`

src/ai/backend/agent/docker/agent.py

@@ -873,6 +873,23 @@ async def start_container(
                             await writer.write(f"{k}={v}\n")
 
                 await container.start()
+
+                if self.internal_data.get("sudo_session_enabled", False):
+                    exec = await container.exec(
+                        [
+                            # file ownership is guaranteed to be set as root:root since command is executed on behalf of root user
+                            "sh",
+                            "-c",
+                            'mkdir -p /etc/sudoers.d && echo "work ALL=(ALL:ALL) NOPASSWD:ALL" > /etc/sudoers.d/01-bai-work',
+                        ],
+                        user="root",
+                    )
+                    shell_response = await exec.start(detach=True)
+                    if shell_response:
+                        raise ContainerCreationError(
+                            container_id=cid,
+                            message=f"sudoers provision failed: {shell_response.decode()}",
+                        )
             except asyncio.CancelledError:
                 if container is not None:
                     raise ContainerCreationError(

src/ai/backend/manager/registry.py

@@ -978,6 +978,9 @@ async def enqueue_session(
         internal_data = {} if internal_data is None else internal_data
         internal_data.update(dotfile_data)
 
+        if sudo_session_enabled:
+            internal_data["sudo_session_enabled"] = True
+
         hook_result = await self.hook_plugin_ctx.dispatch(
             "PRE_ENQUEUE_SESSION",
             (session_id, session_name, access_key),

src/ai/backend/runner/entrypoint.sh

@@ -122,11 +122,6 @@ else
     setsid ssh-add /home/work/.ssh/id_rsa < /dev/null
   fi
 
-  # Enable sudo
-  if [ "$SUDO_SESSION_ENABLED" = "1" ]; then
-    echo "work ALL=(ALL:ALL) NOPASSWD:ALL" >> /etc/sudoers
-  fi
-
   echo "Generate random alpha-numeric password"
   if [ ! -f "$HOME/.password" ]; then
     /opt/kernel/su-exec $USER_ID:$GROUP_ID /opt/backend.ai/bin/python -s /opt/kernel/fantompass.py > "$HOME/.password"