The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes a Python utility (scripts/experiment.py) that executes shell commands to run computational experiments (e.g., training, simulations). This is a core intended functionality.
Evidence: The script uses subprocess.Popen with shell=False to execute commands provided in a structured spec.json file.
Mitigation: The utility performs several security checks: it validates that argv is a list of strings, checks for NUL characters, prevents path traversal to ensure files stay within the designated workspace, and implements resource limits (max runs, wall-time timeouts) to prevent runaway processes. It explicitly avoids using a shell interpreter unless one is provided as the literal executable.
[DATA_EXFILTRATION]: No network connectivity or data exfiltration patterns were detected. The skill's instructions and code explicitly state that it does not connect to the internet, call external APIs, or install external packages.
[PRIVILEGE_ESCALATION]: No privilege escalation attempts were found. The skill operates within the user's local environment and does not request or use sudo or other administrative privileges.
[INDIRECT_PROMPT_INJECTION]: While the skill is designed to ingest and process data from upstream skills (like light-design), it includes instructions to treat external materials as unverified. It uses strictly defined JSON schemas and path validation to limit the attack surface of processed data.

light-experiment