docs: expand GPU Operator security guidance

gpu-operator/security.rst

@@ -1,11 +1,46 @@
 
 *****************************
-Security Considerations
-*****************************
-
-
-Pod Security Context of the Operator and Operands
-=================================================
+Security Considerations
+*****************************
+
+
+Preventing Unprivileged GPU Access
+==================================
+
+By default, a container can request GPU access by setting the
+``NVIDIA_VISIBLE_DEVICES`` environment variable, even when it does not request
+an ``nvidia.com/gpu`` resource. This can allow an unprivileged workload to
+access more GPUs than the Kubernetes device plugin allocated to it.
+
+Cluster administrators can prevent this configuration by making the NVIDIA
+Container Toolkit accept the device list only from the device plugin. Add the
+following values during GPU Operator installation or upgrade:
+
+.. code-block:: yaml
+
+   toolkit:
+     env:
+       - name: ACCEPT_NVIDIA_VISIBLE_DEVICES_ENVVAR_WHEN_UNPRIVILEGED
+         value: "false"
+       - name: ACCEPT_NVIDIA_VISIBLE_DEVICES_AS_VOLUME_MOUNTS
+         value: "true"
+   devicePlugin:
+     env:
+       - name: DEVICE_LIST_STRATEGY
+         value: volume-mounts
+
+This configuration assumes that the cluster's security policy does not allow
+unprivileged workloads to create arbitrary host volume mounts. Review the
+resulting pod security policy and admission controls before enabling it.
+
+Privileged containers can still access all GPUs on a node. Restrict access to
+the Operator namespace and avoid granting privileged access to ordinary
+workloads. Components that need access to all GPUs, such as the device plugin
+and DCGM Exporter, should be treated as infrastructure workloads.
+
+
+Pod Security Context of the Operator and Operands
+=================================================
 
 Several of the NVIDIA GPU Operator operands, such as the driver containers and container toolkit,
 require the following elevated privileges:
@@ -27,8 +62,9 @@ As a best practice, establish proper security policies and prevent any other use
 CVEs
 =================================================
 
-The following is a list of known CVEs in the GPU Operator or its operands.
-To view any published security bulletins for NVIDIA products published security bulletins for NVIDIA products, refer to the NVIDIA product security page at https://www.nvidia.com/en-us/security/.
+The following is a list of known CVEs in the GPU Operator or its operands.
+To view published security bulletins for NVIDIA products, refer to the
+`NVIDIA product security page <https://www.nvidia.com/en-us/security/>`_.
 
 .. list-table:: CVEs
    :widths: 20 45 35
@@ -96,4 +132,5 @@ To view any published security bulletins for NVIDIA products published security
 Report a Vulnerability
 -----------------------------
 
-For details on reporting a suspected vulnerability, refer to the  `GPU Operator Security policies <https://github.com/NVIDIA/gpu-operator/blob/main/SECURITY.md/>`_ page.
+For details on reporting a suspected vulnerability, refer to the `GPU Operator
+Security policy <https://github.com/NVIDIA/gpu-operator/blob/main/SECURITY.md>`_.