The Agent Skills Directory

[COMMAND_EXECUTION]: Several automation scripts, including autoresearch-agent/scripts/run_experiment.py and agenthub/scripts/result_ranker.py, utilize subprocess.run with shell=True to execute shell commands. These commands are typically read from configuration files (like config.cfg) or provided via script arguments. While this is an intended feature for running user-defined benchmarks and test suites, it requires that users trust the source of their project configurations.
[INDIRECT_PROMPT_INJECTION]: The multi-agent collaboration tool agenthub and the optimization loop autoresearch-agent process data from external sources such as git history, shared message boards (markdown files in .agenthub/board/), and structured results files. This creates a vulnerability surface where data generated by one agent or stored in the repository could influence the decisions or prompts of the coordinating agent without rigorous sanitization.
[DYNAMIC_EXECUTION]: The full-page-screenshot tool dynamically generates Python scripts (_stitch.py) at runtime to perform image stitching operations. These scripts are written to the filesystem and then executed using the execSync command.
[REMOTE_CODE_EXECUTION]: The agenthub skill orchestrates parallel sub-agents that operate within isolated git worktrees. These agents are given instructions to modify the codebase and run evaluations. While this follows a structured protocol, the inherent autonomy of these agents and their ability to execute arbitrary code within the repository constitutes a high-capability risk profile.

engineering-advanced-skills