The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains a list of prompt injection strings (e.g., "ignore previous instructions", "DAN", and system prompt override attempts). These are explicitly defined within a detect_injection function and a test-case configuration for defensive monitoring and red-team evaluation. They do not constitute an attempt to hijack the agent's behavior.
[DATA_EXFILTRATION]: Includes a defensive "Canary Token System" and output validation logic to detect and redact sensitive strings like API keys (sk-, AKIA) and canary tokens from model outputs. This is a security-enhancing feature.
[DYNAMIC_EXECUTION]: The provided Python snippets use standard libraries (re, hashlib, json, secrets) to perform input/output sanitization. There is no evidence of eval(), exec(), or other dangerous dynamic execution patterns.
[COMMAND_EXECUTION]: The skill includes a brief example of a shell command (which <tool>) in its documentation. This is used to demonstrate how a developer might check for available tools and is not configured for automated or malicious execution.

prompt-injection-defense