The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run in scripts/run_ab.py to invoke the claude CLI. This is the primary function of the skill (A/B testing prompts). The execution uses a list of arguments and defaults to shell=False, which is the recommended security practice to prevent command injection.- [DATA_EXPOSURE]: The skill captures model responses and execution metrics (tokens, duration, turns) and saves them to a local work directory as JSON files. This data is used only to generate the local HTML report. No network requests or exfiltration attempts were detected.- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted model output to generate HTML reports. The potential risk of Cross-Site Scripting (XSS) is mitigated by consistent use of HTML escaping: html.escape() is used in the Python rendering script, and a custom esc() sanitizer is used in the JavaScript template to handle markdown rendering securely.

instruction-eval