The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides the ability to execute arbitrary shell commands via the command-check and behavior-check verification methods. It uses bash to run commands and bench scripts (e.g., bash bench/run-contract-case.sh) and executes commands provided in the command field of eval cases.
[DYNAMIC_EXECUTION]: The script-check mechanism allows for the execution of custom mcp-scripts defined in project configuration files. These scripts can contain arbitrary logic in languages like JavaScript, which are executed during the evaluation process. Additionally, the skill-check method executes a local Python script (quick_validate.py) with a user-provided target path, which can lead to command injection if the target path string is not properly sanitized.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process data from untrusted or external sources, creating a vulnerability to indirect prompt injection.
Ingestion points: Data enters the skill via the .evals/ directory through files containing pattern-key, promoted-rule, and command strings derived from other automated tools or manual input.
Boundary markers: The skill does not define or enforce boundary markers to separate untrusted rule text from the internal execution logic.
Capability inventory: The skill has high-privilege capabilities including arbitrary shell execution (bash), package manager interaction (npm), and script interpretation (python, javascript).
Sanitization: There is no evidence of sanitization, escaping, or validation of the strings interpolated into commands or scripts before they are executed.

eval-creator