The Agent Skills Directory

[SAFE]: The skill is a set of guidelines and best practices for developers. It does not contain executable malicious code, unauthorized network calls, or hidden instructions.
[PROMPT_INJECTION]: Although the phrase "Ignore previous instructions" appears in SKILL.md, it is used descriptively within a section on 'Response sanitization'. It serves as an example of a threat that developers should defend against when processing external data, rather than an attempt to manipulate the agent's behavior.
[COMMAND_EXECUTION]: The file references/input-hardening.md contains a list of dangerous command strings (e.g., rm -rf /, cat /etc/passwd). These are explicitly identified as fuzzing inputs for a test function (test_input_hardening) designed to ensure that CLI tools correctly reject malicious injection attempts. There is no logic that executes these commands.
[DATA_EXFILTRATION]: The skill promotes safe security practices, such as using environment variables for authentication and implementing path sandboxing to prevent directory traversal. It correctly identifies the risk of agents accessing sensitive files like .ssh keys and provides validation logic to mitigate it.

building-agent-clis