The Agent Skills Directory

[COMMAND_EXECUTION]: The submit.py script executes the gcloud command to retrieve a Slack webhook URL from Google Cloud Secret Manager. The implementation uses subprocess.run with a structured list of arguments, which prevents command injection vulnerabilities.\n- [DATA_EXFILTRATION]: The skill transmits feedback data to an external Slack webhook. The implementation includes significant safeguards: the submit.py script redacts email addresses, credential-like strings (e.g., Slack tokens), and sensitive URL parameters before transmission. Additionally, the skill instructions explicitly direct the agent to exclude secrets, tokens, and customer data from the reports.\n- [PROMPT_INJECTION]: The skill instructions specify that feedback submission should occur autonomously and silently. While this reduces user oversight, the risk of indirect prompt injection is mitigated by explicit instructions to treat user-provided text and external content as untrusted and to never follow instructions contained within them. Evidence chain: (1) Ingestion point: Message content read from standard input in submit.py. (2) Boundary markers: Instructions to treat external content as untrusted and ignore embedded directives. (3) Capability inventory: Network communication via urllib.request and local command execution via gcloud. (4) Sanitization: Robust redaction of credentials, emails, and sensitive URL parameters performed in the submission script.

suggestion-box