Image analysis sandbox

The image analysis sandbox lets you dynamically analyze the runtime behavior of images before running them in your development and production environments.

The analysis mechanism collects and displays container behaviors by safely exercising the image in a sandbox machine. It also exposes risks and identifies suspicious dependencies buried deep in your software supply chain that would otherwise be missed by static analysis for vulnerabilities and compliance issues.

Running the analysis is supported for Linux images on Docker container runtime.

Setup the sandbox machine

In order to run a sandbox analysis for an image, you first need to set up a dedicated sandbox virtual machine.

Prerequisites:

• Install twistcli tool on your machine.
• You need sudo permission to run the twistcli command.
• The sandbox machine should have connectivity to Prisma Cloud Compute Console.
• The machine must be a Linux VM.
• Install Docker on the machine.

When setting up the VM, follow the guidelines below to make sure potential malware doesn’t exploit your sandbox:

• Make sure that the kernel is up to date.
• Make sure that Docker and Runc are up to date.
• Make sure all the software components on the machine are up to date (to make sure there is no other vulnerable component on the machine).
• The VM should be as isolated as possible. Run the VM in a dedicated network, separate from production. If other services run alongside the sandbox VM in the same local network, set up firewall rules to ensure the sandbox VM cannot reach them.
• If the VM runs in the cloud, it shouldn’t run with any service account.

It is recommended to avoid running a Defender on the same machine used as the sandbox VM. Running a Defender on this machine might cause the image that is being analyzed in the sandbox to also be presented under Monitor > Vulnerabilities/Compliance > Images > Deployed images as an image running in the environment.

Running the sandbox command

Description

Triggering a sandbox analysis is done by executing the twistcli sandbox command on an image. After the command is triggered, Prisma Cloud’s sandbox mechanism runs the container, and starts tracing its behavior. The events occurring on the running container are collected and are later analyzed to discover suspicious behaviors.

Synopsis

The usage of the twistcli sandbox command is very similar to running a container image using docker:

$ sudo twistcli sandbox [OPTIONS] IMAGE [COMMAND] [ARG...]

For example:

$ sudo twistcli sandbox --address https://<console-address>:8083 --token 'your-api-token' --analysis-duration 2m -v "$PWD":/app python:3 python3 /app/server.py

To specify an image to scan, use either the image ID, or repository name and tag. The image should be present on the sandbox machine, having either been built or pulled there. If a repository is specified without a tag, twistcli looks for an image tagged latest.

The entrypoint and arguments should be specified after the image. If an entrypoint isn’t specified, the default entrypoint of the image will be used.

Options

--output-file FILENAME Write the results of the analysis to a file in JSON format.

Example: --output-file analysis-results.json

--analysis-duration DURATION The duration of the analysis in a Go duration string format. The default duration is 1 minute.

Adjust the duration according to your image. A longer duration may allow detection of more behaviors. An analysis duration that is too short might cause missing some of the suspicious findings that could have been detected on the container.

Example: --analysis-duration 2m30s

The analysis duration can be shorter than the duration you specified, if the container exits before the analysis time ends.

When WildFire integration is enabled, the analysis duration can be longer than specified, since the communication with WildFire may take longer than the analysis duration. When the specified duration is met, Prisma Cloud stops the container, so no more events are collected, but is waiting for WildFire verdict to publish the results.

-e, --env ENVIRONMENT VAR A key=value pair to define an environment variable in the running container. Repeat flag for each environment variable.

Example: -e "GOROOT=/usr/local/go" -e "HTTPS_PORT=4443"

-v, --volume VOLUME A src:dst pair to mount a volume to the running container. Repeat flag for each mount.

Example: -v "/home/developer/app:/app" -v "/var/lib/mongo:/data"

Any volume that is shared with the sandbox will be accessible to potential malware that exists on the container. Therefore, carefully consider the usage of volumes.

-w, --workdir DIRECTORY Working directory inside the container.

Example: -w "/usr/src/myapp"

--port PORT A host_port:container_port[/tcp|udp] pair to bind a host port the running container’s port. Repeat for each port. Port ranges are not supported.

Example: --port "80:123/tcp"

--third-party-cmd value Specify the third-party script/binary and its arguments

Example: --third-party-cmd /opt/sandbox/openscap_analysis.sh

--third-party-delay value Specify the required time to wait from the container start time (to ensure initialization completion) before executing the third-party command (Optional) (default: "0")

Example: --third-party-delay 5s

--third-party-output value Specify the third party script/binary output path

Example: --third-party-output /opt/sandbox/oscap-results.txt

--tlscacert PATH Path to Prisma Cloud CA certificate file. If no CA certificate is specified, the connection to Console is insecure.

--token TOKEN Token to use for Prisma Cloud Console authentication. Tokens can be retrieved from the API endpoint api/v1/authenticate or from the Manage > System > Utilities page in Console.

--exit-on-error TRUE/FALSE Immediately exit the analysis if an error is encountered.

-h, --help Show help

Return value

The exit code is 0 if the sandbox analysis verdict is "Passed". If the verdict is "Failed", the exit code is 1.

The criteria for passing or failing the sandbox analysis is determined by the severity of the suspicious findings detected during the analysis. The analysis verdict is "Failed" when there is at least one finding with Critical or High severity. Otherwise, the verdict is "Passed".

Another reason why twistcli sandbox might return an exit code of 1 is if the analysis failed due to an error.

Sandbox analysis results

After twistcli dynamically analyzes the image, twistcli:

• Exits with a return value.
• Outputs a summary of the results, including a verdict.
• Outputs a link to the results report in the Console UI.

The results report in the Console UI includes the analysis summary and verdict, a list of suspicious detections found on the image, and the entire container behavior events that occurred during container runtime.

Analysis summary

The analysis summary contains the following main parts:

• Verdict - whether the image passed or failed the analysis.

  The criteria for passing or failing the sandbox analysis is determined by the severity of the suspicious findings detected during the analysis. The analysis verdict is "Failed" when there is at least one finding with Critical or High severity. Otherwise, the verdict is "Passed".
• Highest severity - the severity of the most severe suspicious finding.
• Suspicious findings count - the number of suspicious findings detected.
• Analysis metadata - analysis time, duration, and the container entrypoint.
• Image details - the details of the analyzed image.

  The image details also include an indication of an additional scan that may have been performed on the image. If the image was scanned for vulnerabilities and compliance as a part of the CI process, registry scanning, or as a deployed image, it will be displayed in the Additional scan field. You will also be able to click on its value to see the scan results. Only the furthest stage is reported in the following order: CI → Registry → Deployed.

Suspicious findings

The sandbox analysis mechanism detects the following suspicious behaviors:

Detection	Description	Severity
Malware	Malware detected by WildFire. Detecting malware using WildFire requires the WildFire integration to be enabled. Go to Manage > System > WildFire and turn on the "Enable runtime protection" toggle. You can also choose to upload files with unknown verdicts to WildFire using the matching toggle.	Critical
Crypto miners	Crypto miner was detected.	Critical
Suspicious ELF headers	ELF file with a suspicious header was detected. The binary is either incompatible with the system architecture or the ELF header was manipulated to hinder analysis. For ELF header tampering, Prisma Cloud identifies overlapping headers, deleted headers, and improperly specified section sizes as suspicious.	High
Vertical port scanning	Vertical port scanner was detected.	High
Kernel module modification	Kernel module was being loaded or unloaded.	High
Dropper	A binary that wasn’t included in the original image (dropped on disk) was executed.	High
Modified binary	A process modified a binary.	High
Modified binary execution	Execution of a binary that was included in the original image but has been modified.	High
Fileless Execution	Execution from a memory file descriptor was detected.	High
Fileless executable creation	An executable was written into a memory file descriptor.	High
Executable creation	A new executable file created on the disk.	Medium

Container behavior

The sandbox analysis mechanism collects Processes, Networking, and Filesystem events that occurred while the container was running in the sandbox. The events are displayed in the Console UI analysis report, in order to provide you with an overview of the container behavior at runtime.

There are two display modes for viewing the container behavior events:

• By Type - the events are aggregated by the main event properties, to give you an overview of which process run on the container, what were the network destinations it was trying to reach, what are its listening ports, etc. For example, if a process was running three times, only a single row will appear for this process, with the common properties only (MD5), and without the properties that are changing between events (command, parent process, etc).
• By Time - all the events are presented ordered by the time they occurred. For example, if a process was running three times, three rows with the same process will appear, with different time, and with all the event details for each one of them (command, parent process, etc).

Filesystem events

For container filesystem, Prisma Cloud collects Open, Create, and Modify file events.

Network events

There are three event types collected for container networking:

• Listening port
• Outbound connection
• DNS query

All three types are presented together under the Networking tab, but each has its own properties.

Outbound connection events are also displayed on a world map according to the country matching their IP. Clicking on a connection event will mark it on the map. Hovering a country on the map will show you how many connections were detected for this country.

View sandbox results on image details

When reviewing image details, you can look at its latest sandbox analysis results in a dedicated section. The Anaysis sandbox section contains an analysis summary, including the verdict and the suspicious findings counts by type. Click on the link at the top to move to the full report page.

Actions

Add to trust group

After reviewing the analysis results of an image, you can decide whether you trust this image to run in your development and production environments. Optionally, you can add the image repository to a single or multiple trust groups using the Add to trust group action. This way it is possible for you to get notified or block images that are not trusted. See Trusted Images to learn more.

Export to JSON file

To export the analysis results, use the Export to JSON action at the top of the page. This action will download a file in a JSON format with the analysis results for the image.