The Agent Skills Directory

[COMMAND_EXECUTION]: The main script paper_explain_figures.py uses subprocess.Popen and subprocess.run to manage external processes for image conversion (sips, magick) and model execution (codex, claude). It includes a security check to explicitly forbid the --runner shell option to prevent direct shell command injection.
[INDIRECT_PROMPT_INJECTION]: The skill ingests user notes and snippets of local source code (up to 200KB) and interpolates them into a model prompt. While it attempts to redact secrets, the inclusion of external text without robust delimiters creates an attack surface for indirect prompt injection if the ingested files contain malicious instructions.
[DYNAMIC_EXECUTION]: The local runner mode, intended for testing, constructs a Python script as a string and executes it dynamically using sys.executable -c.
[DATA_EXFILTRATION]: Instructions in SKILL.md guide the agent to use the GitHub CLI tool (gh) to upload bug reports to a repository owned by the author (huangwb8/bensz-bugs) upon explicit user request. This represents a documented path for data to leave the local environment to a vendor-controlled target.
[CREDENTIALS_UNSAFE]: The skill contains defensive regex patterns designed to identify and redact sensitive credentials (e.g., sk-, AKIA, password) from source code snippets before they are transmitted to AI runners. While defensive in nature, these patterns highlight the skill's role in handling potentially sensitive local data.

paper-explain-figures