The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The script scripts/run-ascend-container.sh defaults to a 'privileged' mode that uses the --privileged, --net=host, and --ipc=host Docker flags. These settings bypass standard container isolation, granting the container root-level capabilities on the host system and its network stack.
[DATA_EXFILTRATION]: The script mounts the entire host /home directory to the container (-v /home:/home). This exposure allows any process or image running in the container to access sensitive user data, such as SSH keys, configuration files, and private documents stored on the host.
[COMMAND_EXECUTION]: The skill executes docker run with significant privileges and mounts system-level directories like /usr/local/sbin from the host to the container.
[INDIRECT_PROMPT_INJECTION]: The scripts/run-ascend-container.sh script does not quote the IMAGE and CONTAINER_NAME variables when interpolating them into the shell-executed docker run command. This lack of sanitization could allow for command or flag injection if a user or automated process provides a malicious string as an argument.
Ingestion points: The IMAGE and CONTAINER_NAME arguments in scripts/run-ascend-container.sh (derived from $1 and $2).
Boundary markers: None present.
Capability inventory: The script performs high-privilege docker run operations with device and volume mapping.
Sanitization: None detected for the shell variable interpolation.
[EXTERNAL_DOWNLOADS]: Fetches development images from the official Huawei Ascend Hub (ascendhub.huawei.com).

ascend-docker