The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted manuscript content (.md, .docx, .pdf) through numerous automated auditing scripts. These scripts use regular expressions to extract data (e.g., cohort sizes, author contributions, and reporting markers) which are then aggregated into JSON reports. These reports directly influence the agent's perception of the manuscript's state and may drive automated remediation actions. The lack of explicit boundary markers or sanitization during the ingestion of this external content creates a surface for indirect prompt injection if malicious instructions are embedded within a manuscript file.
[DYNAMIC_EXECUTION]: The script build_marked_manuscript.py generates and executes AppleScript via osascript to automate document comparison in Microsoft Word on macOS. Although the script includes a sanitization helper (_as_literal) to escape quotes and backslashes in interpolated variables like the author's name and file paths, the dynamic construction and execution of scripts based on user or agent-supplied input remains a potential vector for local command execution if the sanitization is bypassed.
[COMMAND_EXECUTION]: The orchestrator script preflight_gate.py uses subprocess.run to execute a suite of internal and external audit tools. While it correctly uses the list-based invocation form to prevent shell injection, it has broad capability to run various Python and Bash scripts across the environment, including scripts located in sibling skill directories.

sync-submission