The Agent Skills Directory

[PROMPT_INJECTION]: The skill defines specific roleplay rules, including start/stop triggers (STOP, EXIT TRIGGER). These instructions are designed to maintain persona consistency and provide a way for users to exit the simulation, rather than attempting to bypass safety guidelines or override core agent instructions.
[DATA_EXPOSURE]: The instructions do not request or hardcode sensitive information such as API keys or private file paths. It emphasizes using publicly available information and web searches to inform the persona's responses.
[INDIRECT_PROMPT_INJECTION]: The skill incorporates external information via a WebSearch tool. It directs the agent to classify questions and conduct research before providing a response. While this involves processing untrusted data from the web, the instructions include boundary checks and a specific "Mental Model" framework to filter and analyze the data before outputting a response, which is a standard functional requirement.
[COMMAND_EXECUTION]: No shell commands, system calls, or unauthorized script executions were identified in the skill instructions or reference materials. The skill strictly uses the platform's standard tool invocation capability.
[OBFUSCATION]: Analysis of the markdown and reference files found no hidden characters, Base64-encoded payloads, homoglyph attacks, or steganographic patterns.

naval-perspective