The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands that interpolate raw user input directly into subshells and variable assignments. Specifically, the command ADVISORS=$("$ADVISOR_SELECTOR" debate standard "$QUESTION") in Step 3 and the prompt construction in Step 5 allow for arbitrary command execution or data exfiltration if the user-provided question contains shell metacharacters like backticks or command substitution syntax.
[INDIRECT_PROMPT_INJECTION]: The skill implements a multi-round debate workflow where the agent is required to read and synthesize responses from external AI advisors. These responses are ingested from the file system (e.g., rounds/r001_*.md) and processed without sanitization or instructions to ignore embedded commands. This creates an attack surface where a malicious or compromised external model could inject instructions to hijack the host agent's session.
[DYNAMIC_EXECUTION]: The instructions provide full bash function definitions (e.g., evaluate_response_quality) and complex conditional logic for the agent to generate and execute via the shell tool at runtime. This dynamic script generation increases the complexity of the execution environment and provides opportunities for users to manipulate the generated code to perform unauthorized actions.

skill-debate