The Agent Skills Directory

[COMMAND_EXECUTION]: The configuration scripts scripts/config_model.py and scripts/config_channel.py are vulnerable to command injection. User-supplied inputs (e.g., API keys, service endpoints, and instance names) are concatenated directly into shell script templates that are subsequently executed with root privileges on remote cloud instances via the Huawei Cloud COC service. This allows for potential unauthorized command execution on the target server if malicious input is provided.
[CREDENTIALS_UNSAFE]: The script scripts/config_model.py stores the user-provided AI model API key in plain text in a local file named model_config_result.json upon completion of the configuration phase, posing a risk of credential exposure to any entity with access to the agent's local filesystem.
[REMOTE_CODE_EXECUTION]: The scripts/install_deps.py script performs remote execution by fetching a setup script from a remote URL (https://deb.nodesource.com/setup_18.x) and piping it directly into a bash shell. NodeSource is a well-known and trusted service for providing Node.js installation scripts.
[DYNAMIC_EXECUTION]: The skill dynamically generates shell scripts at runtime to handle deployment tasks and application configuration. Additionally, scripts/prepare_env.py performs dynamic module loading using __import__ to check for the presence of required Huawei Cloud SDK libraries.
[PRIVILEGE_ESCALATION]: The skill uses root-level execution on remote instances to perform system management tasks, such as installing packages, modifying systemd unit files, and altering global Python environment settings. These high-privilege operations are technically necessary for the deployment goal but represent a significant capability tier.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from user inputs in config_model.py and config_channel.py which is then used to construct commands executed with high privileges. The lack of sanitization at these ingestion points combined with the capability for remote file-writing and code execution identifies a significant attack surface.

huawei-cloud-flexus-l-deploy-jiuwenswarm