The Agent Skills Directory

[COMMAND_EXECUTION]: Several Python helper scripts execute shell commands using subprocess.run with shell=True, incorporating the user-provided software target name without sanitization.
In scripts/doc_generator.py, the get_software_info function executes commands like f"{target} --version" directly in the shell.
In scripts/executor.py, the run_command function executes arbitrary strings retrieved from the installation_plan.yaml file, which is generated based on user input.
This design allows an attacker to execute arbitrary system commands by providing a malicious software name (e.g., using command chaining characters like ; or &).
[REMOTE_CODE_EXECUTION]: The primary function of the skill is to generate and document 'curl piped to bash' installation scripts.
SKILL.md and README.md contain instructions and templates for users to run commands like curl -sSL https://raw.githubusercontent.com/<owner>/<repo>/main/install.sh | bash.
While intended for software distribution, this pattern is a well-known security risk for end-users as it bypasses local verification before execution.
[DATA_EXFILTRATION]: The scripts/env_explorer.py script performs extensive system profiling.
It collects detailed metadata including OS version, CPU architecture, installed package managers, shell paths, user permission levels (root/sudo availability), and system paths (Home, CWD, PATH).
This sensitive environment footprint is saved to env_info.json, which could facilitate targeted attacks if the data is exfiltrated.

install-script-generator