diff --git a/docs/README.md b/docs/README.md
index ed63d60e..9c015a75 100644
--- a/docs/README.md
+++ b/docs/README.md
@@ -58,6 +58,7 @@ User-facing MeTTa skills the agent invokes. Each page follows the template **Sig
 - [reference-channels.md](./reference-channels.md) — IRC, Telegram, Slack, Mattermost, WebSocket, and websearch adapters plus the channel contract
 - [reference-python-bridges.md](./reference-python-bridges.md) — `lib_llm_ext.py`, `src/helper.py`, `src/skills.pl`
 - [reference-memory-portability.md](./reference-memory-portability.md) — Operator backup, restore, and archive-transfer workflow
+- [reference-gpu.md](./reference-gpu.md) - Experimental NVIDIA GPU access: host setup, building and running a GPU-enabled container
 
 ### Plugin API
 
diff --git a/docs/reference-gpu.md b/docs/reference-gpu.md
new file mode 100644
index 00000000..50bf7c8e
--- /dev/null
+++ b/docs/reference-gpu.md
@@ -0,0 +1,107 @@
+# Reference - GPU Support (Experimental)
+
+> **Experimental.** GPU support is under testing and is not included in Omega
+> releases yet; it is planned for a future release. Released images and
+> `scripts/omega start` do not give the agent GPU access. To try it, build the
+> image from the `devices-landlock-fix` branch and start the container with
+> your own `docker run` command, as described below.
+
+## Host setup
+
+### Native Linux
+
+1. Install the NVIDIA driver for your distribution and reboot. `nvidia-smi -L`
+   on the host must list your GPUs.
+2. Install the
+   [NVIDIA Container Toolkit](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html).
+3. Register the toolkit with Docker and restart Docker:
+
+   ```sh
+   sudo nvidia-ctk runtime configure --runtime=docker
+   sudo systemctl restart docker
+   ```
+
+4. Check that CDI is enabled in Docker: `docker info --format '{{json .CDISpecDirs}}'`
+   must print a non-empty list. If it does not, enable it and restart Docker:
+
+   ```sh
+   sudo nvidia-ctk runtime configure --runtime=docker --cdi.enabled=true
+   sudo systemctl restart docker
+   ```
+
+5. Check the CDI specification: `nvidia-ctk cdi list` must print
+   `nvidia.com/gpu=...` entries. Recent toolkit versions keep it up to date with
+   `nvidia-cdi-refresh.service`; otherwise generate it:
+
+   ```sh
+   sudo nvidia-ctk cdi generate --output=/var/run/cdi/nvidia.yaml
+   ```
+
+   Regenerate the specification after every driver update.
+
+### Windows with WSL2
+
+1. Install the NVIDIA driver on Windows. Do not install a Linux NVIDIA driver
+   inside the WSL distribution: WSL exposes the Windows driver as `/dev/dxg`
+   and the libraries in `/usr/lib/wsl/lib`.
+2. Install Docker Engine and the NVIDIA Container Toolkit inside the WSL
+   distribution and follow steps 2-5 of the native Linux setup.
+
+### SELinux (Fedora, RHEL and derivatives)
+
+When Docker runs with SELinux labeling, containers may be denied access to the
+GPU device nodes even though the nodes exist, and `nvidia-smi` fails inside the
+container. Allow containers to use devices:
+
+```sh
+sudo setsebool -P container_use_devices on
+```
+
+### Other setups
+
+- **Rootless Docker** needs extra toolkit configuration, see the rootless mode
+  section of the NVIDIA Container Toolkit install guide.
+- **Docker Desktop** and **Podman** are not covered by these instructions.
+
+## Build the image
+
+Build the image from the `devices-landlock-fix` branch. Its security policy
+allows the agent to use the GPU; images built from other branches do not.
+
+```sh
+git clone https://github.com/singnet/Omega.git
+cd Omega
+git checkout devices-landlock-fix
+docker build -t omega:gpu .
+```
+
+## Start the container
+
+Start the container with your own command and pass the GPUs with
+`--device nvidia.com/gpu=all`, or a single GPU by its CDI name from
+`nvidia-ctk cdi list`, for example `--device nvidia.com/gpu=0`. The other
+options match the ones `scripts/omega` uses. Example for Telegram and OpenAI,
+with `OPENAI_API_KEY`, `TG_BOT_TOKEN` and `OMEGA_AUTH_SECRET` exported in your
+shell:
+
+```sh
+docker rm -f omega
+docker run -d -it \
+  --name omega \
+  --init \
+  --security-opt no-new-privileges:true \
+  --add-host=host.docker.internal:host-gateway \
+  --tmpfs /tmp:size=64m,mode=1777 \
+  --tmpfs /var/tmp:size=64m,mode=1777 \
+  --tmpfs /run:size=16m,mode=755 \
+  --volume omega-memory:/PeTTa/repos/Omega/memory \
+  --device nvidia.com/gpu=all \
+  -e OPENAI_API_KEY \
+  -e TG_BOT_TOKEN \
+  -e OMEGA_AUTH_SECRET \
+  omega:gpu
+```
+
+Do not restart this container with `scripts/omega start`: it recreates the
+container without the GPU options. `docker stop omega` and `docker start omega`
+keep them.
diff --git a/profile/policy.py b/profile/policy.py
index 981718af..eb11ce96 100644
--- a/profile/policy.py
+++ b/profile/policy.py
@@ -71,7 +71,7 @@ class FileSystemPolicy:
                              | AccessFs.REMOVE_DIR | AccessFs.MAKE_FIFO
                              | AccessFs.MAKE_SOCK)
     READ_WRITE_FILE_ACCESS = (AccessFs.READ_FILE | AccessFs.WRITE_FILE |
-                              AccessFs.TRUNCATE)
+                              AccessFs.TRUNCATE | AccessFs.IOCTL_DEV)
 
     def __init__(self):
         self._compatibility = LandLockCompatibility.BEST_EFFORT
@@ -116,11 +116,33 @@ def load_dict(self, policy: dict):
         self._read_only = [Path(f'{p}') for p in ro]
         self._read_write = [Path(f'{p}') for p in rw]
 
+    @staticmethod
+    def _existing_paths(paths: list[Path]) -> list[Path]:
+        """Drop policy paths that do not exist on a host machine.
+
+        Landlock cannot add a rule for a missing path, and the policy lists
+        platform-specific paths such as GPU device nodes that exist only on
+        some host machines.
+
+        Args:
+            paths: Paths listed in the policy.
+
+        Returns:
+            Paths that exist on a host machine.
+        """
+        existing = [p for p in paths if p.exists()]
+        missing = [str(p) for p in paths if not p.exists()]
+        if missing:
+            logger.info(f"Skipped missing policy paths: {missing}")
+        return existing
+
     def apply(self):
-        rod = list(filter(lambda p: p.is_dir(), self._read_only))
-        rof = list(filter(lambda p: not p.is_dir(), self._read_only))
-        rwd = list(filter(lambda p: p.is_dir(), self._read_write))
-        rwf = list(filter(lambda p: not p.is_dir(), self._read_write))
+        ro = self._existing_paths(self._read_only)
+        rw = self._existing_paths(self._read_write)
+        rod = list(filter(lambda p: p.is_dir(), ro))
+        rof = list(filter(lambda p: not p.is_dir(), ro))
+        rwd = list(filter(lambda p: p.is_dir(), rw))
+        rwf = list(filter(lambda p: not p.is_dir(), rw))
 
         strict = self._compatibility == LandLockCompatibility.HARD_REQUIREMENT
         Landlock(strict=strict) \
diff --git a/profile/policy.yaml b/profile/policy.yaml
index 22c3d240..45bd1b4a 100644
--- a/profile/policy.yaml
+++ b/profile/policy.yaml
@@ -5,7 +5,6 @@ filesystem_policy:
   - /bin
   - /usr
   - /lib
-  - /proc
   - /dev/urandom
   - /etc
   - /opt
@@ -21,5 +20,14 @@ filesystem_policy:
   - /opt/sentence_transformers
   - /var/tmp
   - /dev/shm
+  # The CUDA driver names its threads via /proc/self/task/<tid>/comm.
+  - /proc
+  # GPU device nodes: /dev/dxg on WSL2, /dev/nvidia* on native Linux.
+  # Paths missing on the host are skipped.
+  - /dev/dxg
+  - /dev/nvidiactl
+  - /dev/nvidia[0-9]*
+  - /dev/nvidia-uvm
+  - /dev/nvidia-uvm-tools
 landlock:
   compatibility: best_effort