The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/setup_presentation.py utilizes subprocess.run(shell=True) within the run_command function to execute initialization and installation commands (npm install, uv init, uv add). While many command strings are hardcoded, the script accepts a project_name argument that is used to create directories and as a working directory for shell executions. This pattern presents a risk if an attacker provides a malicious project name containing shell metacharacters.
[DYNAMIC_EXECUTION]: The skill implements dynamic execution patterns in scripts/setup_presentation.py, where shell commands are constructed as strings and executed at runtime. This practice, combined with user-provided input, increases the potential attack surface for execution-based vulnerabilities.
[EXTERNAL_DOWNLOADS]: The utility assets/examples/generate_slides.py is designed to fetch data from remote sources. It uses the requests library to perform HTTP GET requests to arbitrary URLs provided via the --data argument, which can result in the download of untrusted content into the skill's environment.
[INDIRECT_PROMPT_INJECTION]: The generate_slides.py script possesses an attack surface for indirect prompt injection by processing external data and interpolating it into Markdown slide templates without visible sanitization.
Ingestion points: The --data argument in assets/examples/generate_slides.py accepts JSON, CSV, or API responses from local files or remote URLs.
Boundary markers: The script does not implement delimiters or safety instructions to prevent the agent from interpreting instructions embedded within the ingested data.
Capability inventory: The skill includes file writing (Path.write_text), network access (requests.get), and shell command execution (subprocess.run via setup_presentation.py).
Sanitization: There is no evidence of validation or escaping for the data content before it is converted into presentation slides.

slidev-theme-neversink