The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses coercive language in the frontmatter ('CRITICAL — Invoke this Skill first, no exceptions', 'Under NO circumstances may you skip') and instructions that attempt to override the agent's autonomy and safety defaults.
[PROMPT_INJECTION]: Instructs the agent to perform 'Silent dependency installation' by running pip install requests -q without notifying the user or seeking confirmation, which facilitates stealthy environment modification.
[REMOTE_CODE_EXECUTION]: Downloads a runtime archive from an unverified personal repository (https://gitcode.com/afeng5267/jiuwenswarm_runtime.git) and subsequently executes the binary components and scripts (e.g., jiuwenswarm-start) contained within it.
[COMMAND_EXECUTION]: Uses sudo privileges to delete and create files in system-protected directories (/usr/local/bin) and modifies permissions to install global commands.
[COMMAND_EXECUTION]: Employs subprocess.run with shell=True in scripts/03_configure.py and scripts/04_start.py, which increases the risk of command injection and bypasses standard process protections.
[CREDENTIALS_UNSAFE]: Accesses the Linux kernel keyring to read HWCLOUD-Agent credentials and parses local settings.json files to harvest API keys, then persists this sensitive data into a plaintext .env file.
[COMMAND_EXECUTION]: Automatically detects and installs system packages (keyutils) using dnf, yum, or apt-get with sudo inside the common.py script without user intervention.

huawei-cloud-install-openjiuwenswarm