The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes Bash(*) for critical file operations. While the primary bash block for directory creation uses proper quoting, the 'Composition rules' section provides instructions for the agent to execute shell commands such as cp -r $PAPER_BASE_DIR/sec/ $NEW_VENUE_DIR/sec/ where path variables are unquoted. This inconsistency creates a risk of command injection or execution errors if the user-provided paths contain spaces or shell metacharacters.\n- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection (Category 8) because it ingests and processes untrusted external data.\n
Ingestion points: The skill reads manuscript content from paper-base-dir and external reviewer feedback from review-corpus.\n
Boundary markers: The instructions lack defined delimiters or clear directives to the agent to disregard malicious commands embedded within these external documents.\n
Capability inventory: The agent has extensive capabilities, including full Bash access and the ability to modify files and invoke other agent skills, which could be exploited if malicious instructions in a reviewer report are processed and executed.\n
Sanitization: No sanitization or validation of the input content is performed before it is used to drive the 'microedit' and 'audit' phases.\n- [SAFE]: The skill demonstrates strong defensive design by creating isolated sibling directories for new submissions, ensuring prior work remains immutable. It also implements a sophisticated '5-layer' anonymity scan to prevent accidental disclosure of author identity and integrates with established academic tools like latexmk, pdfinfo, and /overleaf-sync.

resubmit-pipeline