The Agent Skills Directory

[PROMPT_INJECTION]: The skill ingests untrusted data in the form of raw reviewer comments (Phase 1). There is a risk of indirect prompt injection if a reviewer includes adversarial instructions within their feedback. The instructions do not specify the use of boundary markers or specific sanitization routines to prevent the LLM from executing commands embedded in the review text during the atomization or drafting phases.
[COMMAND_EXECUTION]: The skill is granted access to the Bash(*) tool. While the documentation indicates this is used for robust file writing (heredocs) to prevent write failures, the permission allows for arbitrary shell command execution on the host system.
[EXTERNAL_DOWNLOADS]: The workflow involves querying external bibliographic services like DBLP and CrossRef to verify citations. It also relies on external Model Context Protocol (MCP) servers (Codex and Oracle) to process data and perform stress tests, which involves sending paper and review content to external endpoints.

rebuttal