The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze external review comments pasted by the user from platforms like OpenReview. This represents a potential surface for indirect prompt injection if a reviewer were to include hidden instructions in their feedback. However, the skill's capabilities are confined to local text processing and drafting, and it includes specific automated checks in check_rebuttal.py to identify and flag suspicious content like non-anonymous URLs or identity strings.
[DATA_EXPOSURE]: The skill manages a local .icml/ workspace to store research notes, claims, and draft responses. It includes explicit warnings and instructions for the user to ensure this directory is excluded from final submissions (.gitignore hints) to prevent the accidental disclosure of non-anonymous research data or private notes during the double-blind review process.
[COMMAND_EXECUTION]: The skill utilizes local Python utility scripts (init_workspace.py and check_rebuttal.py) to automate administrative tasks. These scripts rely on standard Python libraries (e.g., pathlib, re, argparse) and perform bounded file system operations within the project directory. No evidence of unauthorized command execution, remote code fetching, or shell injection was found.

icml-rebuttal