Security in SingularityCE
If you suspect you have found a vulnerability in SingularityCE, we want to work with you so that it can be investigated, fixed, and disclosed in a responsible manner. Please follow the steps in our published Security Policy, which begins with contacting us privately via firstname.lastname@example.org.
Sylabs discloses vulnerabilities found in SingularityCE through public CVE reports, as well as notifications on our community channels. We encourage all users to monitor new releases of SingularityCE for security information. Security patches are applied to the latest open-source release.
SingularityPRO is a professionally curated and licensed version of SingularityCE that provides added security, stability, and support beyond that offered by the open-source project. Security and bug-fix patches are backported to select versions of SingularityPRO, so that they can be deployed long-term where required. PRO users receive security fixes as detailed in the Sylabs Security Policy.
SingularityCE grew out of the need to implement a container platform that was suitable for use on shared systems, such as HPC clusters. In these environments, multiple people typically need to access the same shared resource. User accounts, groups, and standard file permissions limit their access to data and devices, and prevent them from disrupting or accessing others’ work.
To provide security in these environments a container needs to run as the user who starts it on the system. Before the widespread adoption of the Linux user namespace, only a privileged user could perform the operations which are needed to run a container. A default Docker installation uses a root-owned daemon to start containers, and users can request that the daemon start a container on their behalf. However, coordinating a daemon with other schedulers is difficult and, since the daemon is privileged, users can ask it to carry out actions that they wouldn’t normally have permission to carry out themselves.
When a user runs a container with SingularityCE, it is started as a normal process running under the user’s account. Standard file permissions and other security controls based on user accounts, groups, and processes apply. In a default installation, SingularityCE uses a setuid starter binary to perform only the specific tasks needed to setup the container.
Setuid & User Namespaces
Using a setuid binary to run container setup operations is essential to support containers on older Linux distributions, such as CentOS 6, that were previously common in HPC and enterprise. Newer distributions have support for ‘unprivileged user namespace creation’. This means a normal user can create a user namespace, in which most setup operations needed to run a container can be run, unprivileged.
SingularityCE supports running containers without setuid, using user
namespaces. It can be compiled with the
--without-setuid option, or
allow setuid = no can be set in
singularity.conf to enable this.
In this mode all operations run as the user who starts the
singularity program. However, there are some disadvantages to this
SIF and other single-file container images cannot be mounted directly. The container image must be extracted to a directory on disk to run. This impacts the speed of execution. As a result, workloads accessing large numbers of small files (as is the case with python application startup, for example) do not benefit from the reduced metadata load on the filesystem an image file provides.
SingularityCE 3.10 introduces experimental functionality to avoid this image-extraction by mounting the SIF container using
squashfuse, if available on your system. You can enable this with the
--sif-fuseflag, or by setting
Replacing direct kernel mounts with a FUSE approach is likely to cause a significant reduction in performance.
The effectiveness of signing and verifying container images is reduced as, when extracted to a directory, modification of files is possible and verification of the image’s original signature cannot be performed.
Encryption is not supported. SingularityCE leverages kernel LUKS2 mounts to run encrypted containers without writing a decrypted version of their content to disk.
Some hold the opinion that vulnerabilities in kernel user namespace code could have greater impact than vulnerabilities confined to a single piece of setuid software, and are therefore reluctant to enable unprivileged user namespace creation.
Because of the points above, the default mode of operation of SingularityCE uses a setuid binary. Sylabs aims to reduce the circumstances that require this as new functionality is developed and reaches commonly deployed Linux distributions.
Runtime & User Privilege Model
While other runtimes have aimed to tackle security concerns by
sandboxing containers executing as the
root user so that they cannot
affect the host system, SingularityCE has adopted a different security
Containers should be run as an unprivileged user.
The user should never be able to elevate their privileges inside the container to gain control over the host.
All permission restrictions on the user outside of a container should apply inside the container, as well.
Favor integration over isolation: a user is allowed to easily use host resources such as GPUs, network filesystems, and high speed interconnects. The process ID space, network, etc., are not isolated in separate namespaces by default.
To accomplish this, SingularityCE uses a number of Linux kernel
features. The container file system is mounted using the
option, and processes are started with the
PR_NO_NEW_PRIVS flag set.
This means that even if you run
sudo inside your container, you
won’t be able to change to another user, or gain root privileges by
If you do require the additional isolation of the network, devices, PIDs, etc., which other runtimes provide, SingularityCE can make use of additional namespaces and functionality such as seccomp and cgroups.
Singularity Image Format (SIF)
SingularityCE uses SIF as its default container format. A SIF container is a single file, which makes it easy to manage and distribute. Inside the SIF file, the container filesystem is held in a SquashFS object. By default, we mount the container filesystem directly using SquashFS. On a network filesystem, this means that reads from the container are data-only. Metadata operations happen locally, speeding up workloads that involve many small files.
Holding the container image in a single file also enables unique security features. The container filesystem is immutable, and can be signed. The signature travels as part of the SIF image itself so that it is always possible to verify that the image has not been tampered with or corrupted.
We use private PGP keys to create a container signature, and the
corresponding public keys to verify the container. Verification of
signed containers happens automatically in
singularity pull commands
against the Sylabs Cloud Container Library. A keystore in the Sylabs
Cloud makes it easier to share and obtain public keys for container
A container could be signed once, by a trusted individual who approves its use. It could also be signed with multiple keys, to signify it has passed each step in a CI/CD QA & Security process. SingularityCE can be configured with an execution control list (ECL), which requires the presence of one or more valid signatures, to limit execution to approved containers.
In SingularityCE 3.4 and above, the root filesystem of a container (stored in the SquashFS partition of SIF) can be encrypted. As a result, everything inside the container becomes inaccessible without the correct key or passphrase. The content of the container then remains private, even if the SIF file is shared in public.
Encryption and decryption are performed using the Linux kernel’s LUKS2 feature. This is the same technology routinely used for full disk encryption. The encrypted container is mounted directly through the kernel. Unlike other container formats, the encrypted container is run without ever decrypting its contents to disk.
Configuration & Runtime Options
System administrators who manage SingularityCE can use configuration files to set security restrictions, grant or revoke a user’s capabilities, manage resources, authorize containers, etc.
For example, the ecl.toml configuration file allows blacklisting and whitelisting of containers.
Documentation for administrators about configuration files and their parameters is available here.
When running a container as root, SingularityCE can apply hardening rules using cgroups, seccomp, and apparmor. See here for details on these options.