@@ -327,6 +327,19 @@ Resource requirements enter the driver layer through `SandboxSpec.resource_requi
 can request a specific number of GPUs or the driver-specific default behaviour.
 For all in-tree drivers, this is equivalent to selecting a single GPU.
 
+For Docker GPU sandboxes, the driver treats CDI specs as runtime metadata for
+both outer injection and inner sandbox policy. It selects opaque CDI device IDs,
+passes them to Docker, embeds a versioned CDI context in the protected workload
+boundary configuration, and mounts daemon-reported CDI spec directories
+read-only into the workload. The sandbox runtime resolves that context in the
+workload mount namespace and derives Landlock paths and supplemental groups
+from CDI `containerEdits` before launching agent processes.
+Host-side CDI spec paths are diagnostic only and are never treated as sandbox
+policy paths.
+Kubernetes must not infer CDI device IDs from the `nvidia.com/gpu` resource
+request; it needs a node-local selected-device handoff before using the same
+workload-side resolver.
+
 VM runtime state paths are derived only from driver-validated sandbox IDs
 matching `[A-Za-z0-9._-]{1,128}`. The gateway-owned VM driver socket uses a
 private `run/` directory plus Unix peer UID/PID checks. Standalone
 
@@ -12,6 +12,12 @@ pub const CDI_CONTEXT_VERSION: u32 = 1;
 /// Base workload-runtime path under which compute drivers mount CDI specification directories.
 pub const CDI_SPEC_DIR_BASE: &str = "/run/openshell/boundary/cdi-specs";
 
+/// Return the workload-runtime path used for a CDI specification directory.
+#[must_use]
+pub fn cdi_spec_mount_path(index: usize) -> String {
+    format!("{CDI_SPEC_DIR_BASE}/{index}")
+}
+
 #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
 pub struct CdiContext {
     pub version: u32,
 
@@ -79,6 +79,7 @@ LSM decisions remain authoritative.
 | Private named volumes | One carries the authenticated sandbox/supervisor channel. The other is mounted only into the supervisor and contains its JWT and private gateway credentials. |
 | In-memory `/run/openshell-supervisor-ca` tmpfs | Holds only the public supervisor CA certificate and trust bundle without making all of `/run` writable. |
 | CDI GPU request | Assigns the exact validated CDI devices requested by driver config or count-based selection. |
+| CDI workload projection | For GPU sandboxes only, embeds selected device metadata in the protected boundary config and mounts Docker daemon CDI spec directories read-only into the workload. |
 
 ## Stop, Start, and Delete
 
@@ -92,6 +93,20 @@ Delete force-removes both containers, the driver-owned runtime volumes, and the
 host-private runtime descriptor. Missing or altered descriptor and channel resources
 fail closed; the driver does not run an older combined-supervisor layout.
 
+## CDI GPU Metadata
+
+Docker remains the source of truth for GPU injection. The driver selects opaque
+CDI device IDs from `driver_config.cdi_devices` or the daemon's discovered CDI
+inventory, then passes the same IDs to Docker with a CDI `DeviceRequest`.
+
+For a GPU sandbox, the driver writes those IDs and projected CDI spec directory
+paths to a versioned context in the protected workload boundary configuration.
+It bind-mounts the daemon-reported spec directories read-only into the workload.
+The sandbox runtime resolves the selected devices in that mount namespace before
+it launches agent processes and derives Landlock paths and supplemental groups
+from CDI `containerEdits`. A missing context or spec directory fails workload
+startup closed. Non-GPU sandboxes receive no CDI context or spec mounts.
+
 ## Driver Config Mounts
 
 The gateway forwards the `docker` block from `--driver-config-json`. Supported
 
@@ -11,6 +11,7 @@ use std::collections::{BTreeMap, HashMap};
 use std::net::IpAddr;
 use std::path::PathBuf;
 
+use openshell_core::cdi::CdiContext;
 use openshell_isolation_interface::contract::{DriverFenceEvidence, ResolvedWorkloadIdentity};
 use openshell_sandbox_backend::GPU_RESOURCE_CLAIM;
 use openshell_sandbox_backend::boundary_protocol::{
@@ -30,6 +31,7 @@ pub struct DockerBoundarySpec {
     pub container_id: String,
     pub image_identity: String,
     pub gpu_requested: bool,
+    pub cdi_context: Option<CdiContext>,
     pub listener_socket: PathBuf,
     pub control_socket: PathBuf,
     pub sandbox_tls: SandboxTlsServerConfig,
@@ -77,7 +79,7 @@ impl DockerBoundarySpec {
                 },
                 resource_claims: resource_claims.clone(),
                 resource_claim_files: BTreeMap::new(),
-                cdi_context: None,
+                cdi_context: self.cdi_context,
                 workload_identity: self.workload_identity.clone(),
                 driver_fence: driver_fence.clone(),
                 child_env: self.child_env,
@@ -123,6 +125,13 @@ mod tests {
             container_id: "sha256:container".to_string(),
             image_identity: "sha256:image".to_string(),
             gpu_requested: true,
+            cdi_context: Some(CdiContext::new(
+                vec!["nvidia.com/gpu=0".to_string()],
+                vec![openshell_core::cdi::CdiSpecDirectory::new(
+                    openshell_core::cdi::cdi_spec_mount_path(0),
+                    "/var/run/cdi",
+                )],
+            )),
             listener_socket: PathBuf::from("/run/openshell/boundary/control.sock"),
             control_socket: PathBuf::from("/host/control.sock"),
             sandbox_tls: SandboxTlsServerConfig {
@@ -158,6 +167,15 @@ mod tests {
             provisioned.runtime_descriptor.resource_claims[GPU_RESOURCE_CLAIM],
             "true"
         );
+        assert_eq!(
+            provisioned
+                .boundary_config
+                .cdi_context
+                .as_ref()
+                .unwrap()
+                .selected_devices,
+            ["nvidia.com/gpu=0"]
+        );
         assert_eq!(
             provisioned.boundary_config.driver_fence,
             provisioned.runtime_descriptor.driver_fence
 
@@ -23,6 +23,7 @@ use bollard::query_parameters::{
 };
 use bytes::Bytes;
 use futures::{Stream, StreamExt};
+use openshell_core::cdi::{CdiContext, CdiSpecDirectory, cdi_spec_mount_path};
 use openshell_core::config::DEFAULT_STOP_TIMEOUT_SECS;
 use openshell_core::driver_mounts;
 use openshell_core::driver_utils::{
@@ -319,12 +320,50 @@ struct DockerDriverRuntimeConfig {
     app_armor_profile: Option<AppArmorProfile>,
 }
 
-#[derive(Debug, Clone, Copy)]
+#[derive(Debug, Clone)]
 struct DockerGpuRuntimeCapabilities {
     cdi_supported: bool,
+    cdi_spec_dirs: Vec<String>,
     wsl_all_gpu_fallback_enabled: bool,
 }
 
+impl DockerGpuRuntimeCapabilities {
+    fn cdi_context(&self, gpu_device_ids: Option<&[String]>) -> Result<Option<CdiContext>, Status> {
+        let Some(gpu_device_ids) = gpu_device_ids.filter(|device_ids| !device_ids.is_empty())
+        else {
+            return Ok(None);
+        };
+        self.require_cdi_spec_dirs()?;
+        Ok(Some(CdiContext::new(
+            gpu_device_ids.to_vec(),
+            self.cdi_spec_dirs
+                .iter()
+                .enumerate()
+                .map(|(index, source)| CdiSpecDirectory::new(cdi_spec_mount_path(index), source))
+                .collect(),
+        )))
+    }
+
+    fn cdi_spec_bind_strings(&self) -> Result<Vec<String>, Status> {
+        self.require_cdi_spec_dirs()?;
+        Ok(self
+            .cdi_spec_dirs
+            .iter()
+            .enumerate()
+            .map(|(index, source)| format!("{source}:{}:ro,z", cdi_spec_mount_path(index)))
+            .collect())
+    }
+
+    fn require_cdi_spec_dirs(&self) -> Result<(), Status> {
+        if self.cdi_spec_dirs.is_empty() {
+            return Err(Status::failed_precondition(
+                "docker GPU sandboxes require Docker CDI spec directories reported by the daemon",
+            ));
+        }
+        Ok(())
+    }
+}
+
 #[derive(Debug, Clone, PartialEq, Eq)]
 enum DockerGatewayRoute {
     Bridge { bind_address: SocketAddr },
@@ -864,14 +903,13 @@ impl DockerComputeDriver {
         let info = docker.info().await.map_err(|err| {
             Error::execution(format!("failed to query Docker daemon info: {err}"))
         })?;
-        let cdi_supported = info
-            .cdi_spec_dirs
-            .as_ref()
-            .is_some_and(|dirs| !dirs.is_empty());
+        let cdi_spec_dirs = info.cdi_spec_dirs.clone().unwrap_or_default();
+        let cdi_supported = !cdi_spec_dirs.is_empty();
         let cdi_gpu_inventory = docker_cdi_gpu_inventory(&info);
         let wsl_all_gpu_fallback_enabled = docker_info_reports_wsl2(&info);
         let gpu = DockerGpuRuntimeCapabilities {
             cdi_supported,
+            cdi_spec_dirs,
             wsl_all_gpu_fallback_enabled,
         };
         validate_sandbox_pids_limit(docker_config.sandbox_pids_limit)?;
@@ -973,7 +1011,7 @@ impl DockerComputeDriver {
                 ssh_socket_path: docker_config.ssh_socket_path.clone(),
                 guest_tls,
                 daemon_version: version.version.unwrap_or_else(|| "unknown".to_string()),
-                gpu,
+                gpu: gpu.clone(),
                 sandbox_pids_limit: docker_config.sandbox_pids_limit,
                 enable_bind_mounts: docker_config.enable_bind_mounts,
                 upstream_proxy: docker_config.upstream_proxy.clone(),
@@ -1616,7 +1654,7 @@ impl DockerComputeDriver {
             &created.id,
             &image,
             &workload_identity,
-            gpu_devices.is_some(),
+            gpu_devices.as_deref(),
         )
         .await
         {
@@ -4387,7 +4425,7 @@ async fn prepare_docker_boundary_files(
     container_id: &str,
     image: &DockerImageMetadata,
     workload_identity: &ResolvedWorkloadIdentity,
-    gpu_requested: bool,
+    gpu_device_ids: Option<&[String]>,
 ) -> Result<(), Status> {
     let directory = docker_boundary_state_dir(sandbox, config)?;
     let workspace_root = driver_mounts::resolve_oci_workspace_root(&image.working_dir)
@@ -4418,7 +4456,8 @@ async fn prepare_docker_boundary_files(
         verification_keys,
         container_id: container_id.to_string(),
         image_identity: image.id.clone(),
-        gpu_requested,
+        gpu_requested: gpu_device_ids.is_some_and(|device_ids| !device_ids.is_empty()),
+        cdi_context: config.gpu.cdi_context(gpu_device_ids)?,
         listener_socket: PathBuf::from(BOUNDARY_SOCKET_MOUNT_PATH),
         control_socket: PathBuf::from(BOUNDARY_SOCKET_MOUNT_PATH),
         sandbox_tls: SandboxTlsServerConfig {
@@ -4996,6 +5035,7 @@ async fn spawn_docker_control_process(
             // fenced by network=none.
             network_mode: Some(config.network_name.clone()),
             mounts: Some(supervisor_mounts),
+            binds: None,
             cap_drop: Some(vec!["ALL".to_string()]),
             cap_add: None,
             security_opt: Some(vec!["no-new-privileges:true".to_string()]),
@@ -5543,7 +5583,10 @@ fn build_container_create_body_for_image(
         }),
         ..Default::default()
     });
-    let user_bind_strings = docker_driver_bind_strings(driver_config)?;
+    let mut user_bind_strings = docker_driver_bind_strings(driver_config)?;
+    if gpu_device_ids.is_some_and(|device_ids| !device_ids.is_empty()) {
+        user_bind_strings.extend(config.gpu.cdi_spec_bind_strings()?);
+    }
     let device_requests = gpu_device_ids.map(|device_ids| {
         vec![DeviceRequest {
             driver: Some("cdi".to_string()),
 
@@ -26,6 +26,9 @@ use std::net::{IpAddr, Ipv4Addr, SocketAddr};
 use std::sync::Arc;
 use tempfile::TempDir;
 
+const TEST_CDI_SPEC_DIR: &str = "/opt/openshell-test/cdi";
+const TEST_CDI_SPEC_DIR_ALT: &str = "/srv/openshell-test/cdi";
+
 fn test_launch_authentication() -> Vec<u8> {
     serde_json::to_vec(&SandboxLaunchAuthentication {
         supervisor: SupervisorAuthBundle {
@@ -150,6 +153,7 @@ fn runtime_config() -> DockerDriverRuntimeConfig {
         daemon_version: "28.0.0".to_string(),
         gpu: DockerGpuRuntimeCapabilities {
             cdi_supported: false,
+            cdi_spec_dirs: Vec::new(),
             wsl_all_gpu_fallback_enabled: false,
         },
         sandbox_pids_limit: openshell_core::config::default_sandbox_pids_limit(),
@@ -160,6 +164,59 @@ fn runtime_config() -> DockerDriverRuntimeConfig {
     }
 }
 
+fn runtime_config_with_cdi_spec_dirs(cdi_spec_dirs: &[&str]) -> DockerDriverRuntimeConfig {
+    let mut config = runtime_config();
+    config.gpu.cdi_supported = !cdi_spec_dirs.is_empty();
+    config.gpu.cdi_spec_dirs = cdi_spec_dirs
+        .iter()
+        .map(|path| (*path).to_string())
+        .collect();
+    config
+}
+
+#[test]
+fn gpu_runtime_builds_cdi_context_from_selected_devices_and_daemon_dirs() {
+    let config = runtime_config_with_cdi_spec_dirs(&[TEST_CDI_SPEC_DIR, TEST_CDI_SPEC_DIR_ALT]);
+    let selected_devices = vec![
+        "nvidia.com/gpu=0".to_string(),
+        "nvidia.com/gpu=1".to_string(),
+    ];
+
+    let context = config
+        .gpu
+        .cdi_context(Some(&selected_devices))
+        .unwrap()
+        .expect("GPU selection should produce a CDI context");
+
+    assert_eq!(context.selected_devices, selected_devices);
+    assert_eq!(
+        context.spec_dirs,
+        vec![
+            CdiSpecDirectory::new(cdi_spec_mount_path(0), TEST_CDI_SPEC_DIR),
+            CdiSpecDirectory::new(cdi_spec_mount_path(1), TEST_CDI_SPEC_DIR_ALT),
+        ]
+    );
+}
+
+#[test]
+fn gpu_runtime_omits_cdi_projection_without_selected_devices() {
+    let config = runtime_config_with_cdi_spec_dirs(&[TEST_CDI_SPEC_DIR]);
+    assert_eq!(config.gpu.cdi_context(None).unwrap(), None);
+}
+
+#[test]
+fn gpu_runtime_builds_read_only_cdi_bind_mounts() {
+    let config = runtime_config_with_cdi_spec_dirs(&[TEST_CDI_SPEC_DIR, TEST_CDI_SPEC_DIR_ALT]);
+
+    assert_eq!(
+        config.gpu.cdi_spec_bind_strings().unwrap(),
+        vec![
+            format!("{TEST_CDI_SPEC_DIR}:{}:ro,z", cdi_spec_mount_path(0)),
+            format!("{TEST_CDI_SPEC_DIR_ALT}:{}:ro,z", cdi_spec_mount_path(1)),
+        ]
+    );
+}
+
 fn test_workload_identity() -> ResolvedWorkloadIdentity {
     ResolvedWorkloadIdentity::new(
         1234,
@@ -2638,8 +2695,7 @@ fn validate_sandbox_auth_accepts_launch_authentication() {
 
 #[test]
 fn build_container_create_body_maps_default_gpu_to_selected_cdi_device() {
-    let mut config = runtime_config();
-    config.gpu.cdi_supported = true;
+    let config = runtime_config_with_cdi_spec_dirs(&[TEST_CDI_SPEC_DIR]);
     let mut sandbox = test_sandbox();
     sandbox.spec.as_mut().unwrap().resource_requirements = Some(gpu_resources(None));
 
@@ -2664,6 +2720,15 @@ fn build_container_create_body_maps_default_gpu_to_selected_cdi_device() {
         request.device_ids.as_ref().unwrap(),
         &vec!["nvidia.com/gpu=1".to_string()]
     );
+    let binds = create_body
+        .host_config
+        .as_ref()
+        .and_then(|host_config| host_config.binds.as_ref())
+        .expect("GPU request should project CDI specs into the workload");
+    assert!(binds.contains(&format!(
+        "{TEST_CDI_SPEC_DIR}:{}:ro,z",
+        cdi_spec_mount_path(0)
+    )));
 }
 
 #[test]
@@ -2686,8 +2751,7 @@ fn build_container_create_body_omits_devices_without_resolved_default_cdi_device
 
 #[test]
 fn build_container_create_body_passes_explicit_cdi_device_id_through() {
-    let mut config = runtime_config();
-    config.gpu.cdi_supported = true;
+    let config = runtime_config_with_cdi_spec_dirs(&[TEST_CDI_SPEC_DIR]);
     let mut sandbox = test_sandbox();
     let spec = sandbox.spec.as_mut().unwrap();
     spec.resource_requirements = Some(gpu_resources(None));
@@ -2757,8 +2821,7 @@ fn build_container_create_body_rejects_empty_cdi_devices() {
 
 #[test]
 fn driver_default_gpu_selection_consumes_distinct_devices_for_creates() {
-    let mut config = runtime_config();
-    config.gpu.cdi_supported = true;
+    let config = runtime_config_with_cdi_spec_dirs(&[TEST_CDI_SPEC_DIR]);
     let driver = test_driver_with_config(config);
     driver.gpu_selector.refresh(
         CdiGpuInventory::new(["nvidia.com/gpu=0", "nvidia.com/gpu=1"]),
 
@@ -196,6 +196,17 @@ idempotent start request. Explicitly stopped sandboxes remain stopped.
 
 For GPU-backed Docker sandboxes, configure Docker CDI before starting the gateway so OpenShell can detect the daemon capability.
 
+For Docker GPU/CDI sandboxes, OpenShell uses Docker's selected CDI device IDs
+and daemon-reported CDI spec directories to build a workload CDI context. The
+driver embeds the context in the protected boundary configuration and
+bind-mounts the spec directories read-only into the workload. If context
+creation fails, the driver removes the created state; if workload or supervisor
+startup fails, it also removes the runtime resources before reporting the
+failure. The sandbox runtime derives filesystem and supplemental group
+requirements from CDI specs before launching agent processes. Non-GPU Docker
+sandboxes do not receive the CDI context, spec mounts, or CDI-derived policy
+changes.
+
 ### Docker Driver Config Mounts
 
 Docker driver config accepts user-supplied `volume` and `tmpfs` mounts. It also