The Agent Skills Directory

[REMOTE_CODE_EXECUTION] (CRITICAL): The skill documentation explicitly describes a 'Schema Mechanism' where natural language conditions are compiled into executable JS/PY code (e.g., complete_when_js: "(ctx) => ctx.player.inventory.includes('treasure')"). This design pattern promotes the dynamic evaluation of potentially untrusted strings as code, a severe security risk.
[COMMAND_EXECUTION] (HIGH): The skill is granted read_file and write_file permissions. When coupled with the dynamic execution logic described in the documentation, an attacker who successfully injects a malicious 'Goal' definition could gain unauthorized access to the underlying file system.
[PROMPT_INJECTION] (HIGH): As a skill meant to process quest objectives (likely from external or untrusted narrative data), it lacks boundary markers or sanitization logic. The complete_when and reward fields represent an injection surface where malicious instructions could be embedded to manipulate the agent's logic or downstream actions.

goal