gsc – Graphene Shielded Containers


GSC is still under development and must not be used in production! Please see issue #1520 for a description of missing features and security caveats.




Docker containers are widely used to deploy applications in the cloud. Using Graphene Shielded Containers (GSC) we provide the infrastructure to deploy Docker containers protected by Intel SGX enclaves using the Graphene Library OS.

The gsc tool transforms a Docker image into a new image (called gsc-<image-name>) which includes the Graphene Library OS, manifest files, Intel SGX related information, and executes the application inside an Intel SGX enclave using the Graphene Library OS. It follows the common Docker approach to first build an image and subsequently run a container of an image. At first a Docker image has to be graphenized via the gsc build command. When the graphenized image should run within an Intel SGX enclave, the image has to be signed via a gsc sign-image command. Subsequently, the image can be run using docker run.


The installation descriptions of prerequisites are for Ubuntu 18.04 and may differ when using a different Ubuntu version or Linux distribution.

Software packages

Please install the, python3, python3-pip packages. In addition, install the Docker client, Jinja2, TOML, and YAML python packages via pip. GSC requires Python 3.6 or later.

sudo apt-get install python3 python3-pip
pip3 install docker jinja2 toml pyyaml

SGX software stack

To run with Intel SGX, please install the corresponding software stack as described in Building.

Host configuration

To create Docker images, the user must have access to Docker daemon.


Please use this step with caution. By granting the user access to the Docker group, the user may acquire root privileges via docker run.

sudo adduser $USER docker

Create a configuration file called config.yaml or specify a different configuration file via gsc option. Please see the documentation on configuration options below and use the config.yaml.template as reference.

Command line arguments


Display usage.

gsc build – build graphenized image

Builds an unsigned graphenized Docker image of an application image called gsc-<IMAGE-NAME>-unsigned by compiling Graphene or relying on a prebuilt Graphene image.



Compile Graphene with debug flags and debug output. If configured to use a prebuilt Graphene image, the image has to support this option.


Compile Graphene with Linux PAL in addition to Linux-SGX PAL. If configured to use a prebuilt Graphene image, the image has to support this option.


Allow untrusted arguments to be specified at docker run. Otherwise any arguments specified during docker run are ignored.


Disable Docker’s caches during gsc build. This builds the unsigned graphenized image from scratch.


Remove intermediate Docker images created by gsc build, if the image build is successful.


Set build-time variables during gsc build (same as docker build –build-arg).


Specify configuration file. Default: config.yaml.


Name of the application Docker image.


Manifest file (Graphene configuration).

gsc sign-image – signs a graphenized image

Signs the enclave of an unsigned graphenized Docker image and creates a new Docker image called gsc-<IMAGE-NAME>. gsc sign-image always removes intermediate Docker images, if successful or not, to ensure the removal of the signing key in them.

gsc sign-image [OPTIONS] <IMAGE-NAME> <KEY-FILE>


Specify configuration file. Default: config.yaml


Name of the application Docker image


Used to sign the Intel SGX enclave

gsc build-graphene – build Graphene-only Docker image

Builds a base Docker image including the Graphene sources and compiled runtime. This base image can be used as input for gsc build via configuration parameter Graphene.Image.

gsc build-graphene [OPTIONS] <IMAGE-NAME>


Compile Graphene with debug flags and debug output. Allows gsc build commands to include debug runtime using -d.


Compile Graphene with Linux PAL in addition to Linux-SGX PAL. Allows gsc build commands to include the Linux PAL using -L.


Disable Docker’s caches during gsc build-graphene. This builds the unsigned graphenized image from scratch.


Remove intermediate Docker images created by gsc build-graphene, if the image build is successful.


Set build-time variables during gsc build-graphene (same as docker build –build-arg).


Specify configuration file. Default: config.yaml


Stop after Dockerfile is created and do not build the Docker image.


Name of the resulting Graphene Docker image

gsc info-image – retrieve information about graphenized Docker image

Retrieves Intel SGX relevant information about the graphenized Docker image such as the MRENCLAVE and MRSIGNER measurements for each application in the Docker image.


gsc info-image <IMAGE-NAME>


Name of the graphenized Docker image

Using Graphene’s trusted command line arguments

Most executables aren’t designed to run with attacker-controlled arguments. Allowing an attacker to control executable arguments can break the security of the resulting enclave.

gsc build uses the existing Docker image’s entrypoint and cmd fields to identify the trusted arguments. These arguments are stored in trusted_argv. This file is only generated when --insecure-args is not specified. As a result any arguments specified during docker run are ignored.

To be able to provide arguments at runtime, the image build has to enable this via the option --insecure-args.

Stages of building graphenized SGX Docker images

The build process of a graphenized Docker image from image <image-name> follows three main stages and produces an image named gsc-<image-name>. gsc build-graphene performs only the first stage, gsc build performs the first two stages, and finally gsc sign-image performs the last stage.

  1. Building Graphene. The first stage builds Graphene from sources based on the provided configuration (see config.yaml) which includes the distribution (e.g., Ubuntu 18.04), Graphene repository, and the Intel SGX driver details. This stage can be skipped if gsc build uses a pre-built Graphene Docker image.
  2. Graphenizing the application image. The second stage copies the important Graphene artifacts (e.g., the runtime and signer tool) from the first stage (or if the first stage was skipped, it pulls a prebuilt Docker image defined via the configuration file). It then prepares image-specific variables such as the executable path and the library path, and scans the entire image to generate a list of trusted files. GSC excludes files and paths starting with /boot, /dev, .dockerenv, .dockerinit, /etc/mtab, /etc/rc, /proc, /sys, and /var, since checksums are required which either don’t exist or may vary across different deployment machines. GSC combines these variables and list of trusted files into a new manifest file. In a last step the entrypoint is changed to launch the script which generates an Intel SGX token and starts the graphene-sgx loader. Note that the generated image (gsc-<image-name>-unsigned) cannot successfully load an Intel SGX enclave, since essential files and the signature of the enclave are still missing (see next stage).
  3. Signing the Intel SGX enclave. The third stage uses Graphene’s signer tool to generate SIGSTRUCT files for SGX enclave initialization. This tool also generates an SGX-specific manifest file. The required signing key is provided by the user via the gsc sign-image command and copied into this Docker build stage. The generated image is called gsc-<image-name> and includes all necessary files to start an Intel SGX enclave.

In the future we plan to provide prebuilt Graphene images for popular cloud-provider offerings.

Generating a signed graphenized Docker image

The last stage combines the graphenized Docker image with the signed enclave and manifest files. Therefore it copies the SIGSTRUCT files and the SGX-specific manifest file from the previous stage into the graphenized Docker image from the second stage.


GSC is configured via a configuration file called config.yaml or specified as a gsc option. A template configuration file is provided in config.yaml.template.


Defines Linux distribution to be used to build Graphene in. Currently the only supported value is ubuntu18.04.


Source repository of Graphene. Default value:


Use this branch of the repository. Default value: master.


Builds graphenized Docker image based on a prebuilt Graphene Docker image. These images are prepared via gsc build-graphene and will be provided for popular cloud-provider environments. Graphene.Repository and Graphene.Branch are ignored in case Graphene.Image is specified.


Source repository of the Intel SGX driver. Default value: “” (in-kernel driver)


Use this branch of the repository. Default value: “” (in-kernel driver)

Run graphenized Docker images

Execute docker run command via Docker CLI and provide gsgx and isgx/sgx devices and the PSW/AESM socket. Additional Docker options and executable arguments may be supplied to the docker run command.


Forwarding devices to a container lowers security of the host. GSC should never be used as a sandbox for applications (i.e. it only shields the app from the host but not vice versa).

docker run [OPTIONS] gsc-<IMAGE-NAME> [<ARGUMENTS>]


docker run options. Common options include -it (interactive with terminal), -d (detached), --device (forward device). Please see Docker manual for details.


Name of original image (without GSC build).


Arguments to be supplied to the executable launching inside the Docker container and Graphene. Such arguments may only be provided when --insecure-args was specified during gsc build.

Execute with Linux PAL instead of Linux-SGX PAL

When specifying -L during GSC gsc build, you may select the Linux PAL at Docker run time instead of the Linux-SGX PAL by specifying the environment variable GSC_PAL as an option to the docker run command. When using the Linux PAL, it is not necessary to sign the image via a gsc sign-image command.


This environment variable specifies the pal loader.

docker run ... --env GSC_PAL=Linux gsc-<image-name> ...


The test folder in Tools/gsc describes how to graphenize Docker images and test them with sample inputs. The samples include Ubuntu-based Docker images of Bash, Python, Node.js, Numpy, Pytorch, and a few more.


All test images rely on insecure arguments to be able to set test-specific arguments to each application. These images are not intended for production environments.

The example below shows how to graphenize the public Docker image of Python3. This example assumes that all prerequisites are installed and configured.

  1. Create a configuration file:

    cd Tools/gsc
    cp config.yaml.template config.yaml
    # Manually adopt config.yaml to the installed Intel SGX driver and desired
    # Graphene repository/version.
  2. Generate the signing key (if you don’t already have a key):

    openssl genrsa -3 -out enclave-key.pem 3072
  3. Pull public Python image from Dockerhub:

    docker pull python
  4. Graphenize the Python image using gsc build:

    ./gsc build --insecure-args python test/ubuntu18.04-python3.manifest
  5. Sign the graphenized Docker image using gsc sign-image:

    ./gsc sign-image python enclave-key.pem
  6. Retrieve SGX-related information from graphenized image using gsc info-image:

    ./gsc info-image gsc-python
  7. Test the graphenized Docker image (change --device=/dev/isgx to your version of the Intel SGX driver if needed):

    docker run --device=/dev/gsgx --device=/dev/isgx \
       -v /var/run/aesmd/aesm.socket:/var/run/aesmd/aesm.socket \
       gsc-python -c 'print("HelloWorld!")'
  8. You can also start a Bash interactive session in the graphenized Docker image (useful for debugging):

    docker run --device=/dev/gsgx --device=/dev/isgx \
       -v /var/run/aesmd/aesm.socket:/var/run/aesmd/aesm.socket \
       -it --entrypoint /bin/bash gsc-python


This document focuses on the most important limitations of GSC. Issue #1520 provides the complete list of known limitations and serves as a discussion board for workarounds.

Dependency on Ubuntu 18.04

Docker images not based on Ubuntu 18.04 may not be compatible with GSC. GSC relies on Graphene to execute Linux applications inside Intel SGX enclaves and the installation of prerequisites depends on package manager and package repositories. GSC can simply be extended to support other distributions by providing a template for this distribution in Tools/gsc/templates.

Trusted data in Docker volumes

Data mounted as Docker volumes at runtime is not included in the general search for trusted files during the image build. As a result, Graphene denies access to these files, since they are neither allowed nor trusted files. This will likely break applications using files stored in Docker volumes.


Trusted files can be added to image-specific manifest file (first argument to gsc build command) at build time. This workaround does not allow these files to change between build and run, or over multiple runs. This only provides integrity for files and not confidentiality.

Allowing dynamic file contents via Graphene protected files

Docker volumes can include Graphene protected files. As a result Graphene can open these protected files without knowing the exact contents as long as the protected file was configured in the manifest. The complete and secure use of protected files may require additional steps.

Integration of Docker Secrets

Docker Secrets are automatically pulled by Docker and the results are stored either in environment variables or mounted as files. GSC is currently unaware of such files and hence, cannot mark them trusted. Similar to trusted data, these files may be added to the manifest.

Access to files in excluded paths

The manifest generation excludes all files and paths starting with /boot , /dev, .dockerenv, .dockerinit, /etc/mtab, /etc/rc, /proc, /sys, and /var from the list of trusted files. If your application relies on some files in these directories, you must manually add them to the manifest:

sgx.trusted_files.[identifier] = "[URI]"
sgx.allowed_files.[identifier] = "[URI]"

Docker images with non-executables as entrypoint

Docker images may contain a script entrypoint which is not an ELF executable. gsc fails to recognize such entrypoints and fails during the image build. A workaround relies on creating an image from the application image which has an entrypoint of the script interpreter with the script as an argument. This allows gsc to start the interpreter instead of the script.