The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from user-provided papers (PDF/Word documents) which can contain malicious instructions intended to manipulate the agent.
Ingestion points: The agent reads the full content of user-provided papers in SKILL.md Step 1 using the platform's Read tool or the pdftotext utility.
Boundary markers: There are no instructions in the prompt to use delimiters or to treat the ingested paper content as untrusted text that should not be followed as instructions.
Capability inventory: The skill can read local files, execute shell commands (pdftotext), and write new Excel files to the user's file system via a Python script.
Sanitization: No sanitization, escaping, or filtering of the extracted text is performed before it is processed by the agent.
[COMMAND_EXECUTION]: The skill instructions involve executing external commands and local scripts.
The agent is directed to use the pdftotext shell utility to extract text from PDF files if direct reading fails.
The skill executes a local script scripts/export_excel.py via subprocess.run to generate Excel reports. While the provided Python implementation uses list-based arguments (a safe practice), the use of external subprocesses increases the overall attack surface if input is not carefully handled by the agent.

paper-analyzer