The Agent Skills Directory

[Assessment Framework for AI Security]: The skill provides a detailed methodology for conducting 'red team' or adversarial testing of AI agents. It guides the user through mapping attack surfaces, generating test cases for indirect prompt injection and oversharing, and scoring findings based on impact. This is a standard security assurance practice.
[Authorization and Safety Guardrails]: A significant portion of the instructions is dedicated to an 'Authorization gate' and 'Guardrails.' These require the user to confirm they own the agent or have written permission, use only synthetic 'canary' data instead of real sensitive information, and avoid destructive actions on live systems. These measures are designed to prevent the skill from being used for unauthorized or harmful activities.
[Indirect Prompt Injection Assessment]: The skill explicitly teaches how to identify vulnerabilities to indirect prompt injection (where instructions are hidden in data the agent retrieves). It classifies this as a high-priority risk factor and provides methods to test for it using synthetic tokens.
[Local Script Execution]: The skill includes a Python script (scripts/canaries.py) intended to be run locally by the user. The script generates CSV files and a markdown checklist to track the progress of the security review. It uses only standard libraries and does not perform network operations, which is consistent with its stated purpose of local data scaffolding.

agent-red-team