The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill explicitly suggests using sshpass -p 'password' for non-interactive remote server access. Passing passwords as command-line arguments is an insecure practice that exposes credentials to other users on the system via the process list and command history.- [PRIVILEGE_ESCALATION]: The containerized deployment script (scripts/docker-commands.sh) instructs the agent to create Docker containers using the --privileged=true flag. Additionally, it mounts sensitive host file system paths into the container, such as /etc/ascend_install.info, /usr/local/sbin, and /usr/local/Ascend/driver, which significantly increases the risk of a container breakout and compromise of the host server.- [INDIRECT_PROMPT_INJECTION]: The workflow involves automatically extracting and executing installation commands from a README.md file located in a user-controlled path ({work_path}/msmodelslim/example/{model_name}/README.md). This creates a vulnerability where a malicious repository could contain instructions that trick the agent into installing harmful packages.
Ingestion points: User-provided work_path and model_name variables determine the target files for data extraction in scripts/env-check.sh and templates/env-checklist.md.
Boundary markers: There are no explicit markers or safety instructions to distinguish between legitimate documentation and embedded malicious instructions in the parsed README files.
Capability inventory: The skill has extensive permissions, including the ability to run shell scripts (bash install.sh), install Python packages via pip, and manage Docker containers.
Sanitization: No verification, filtering, or escaping is performed on the commands extracted from the external documentation files before they are executed in the shell.- [EXTERNAL_DOWNLOADS]: The skill fetches the msmodelslim toolkit from the official Ascend GitCode repository (https://gitcode.com/Ascend/msmodelslim.git) and executes a shell script (bash install.sh) to perform the setup. This behavior is documented as part of the core installation process for the vendor's tool.- [COMMAND_EXECUTION]: The skill employs various shell-based tools for environment validation, NPU status monitoring, and remote execution, including npu-smi, docker, and sshpass.

msmodelslim-quant