The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill represents a significant surface for indirect prompt injection attacks.
Ingestion points: The skill downloads academic papers (PDFs) and fetches web content which are processed in scripts/pflib/verify.py (using pdftotext) and scripts/pflib/browser.py (via content extraction).
Boundary markers: While SKILL.md provides detailed briefs for subagents (paper agent and judge), it lacks explicit delimiters or instructions to ignore embedded commands within the extracted paper content.
Capability inventory: The skill possesses extensive capabilities, including shell command execution (subprocess.run, os.execvp) and network operations (urllib, opencli).
Sanitization: No specific sanitization or filtering of the extracted text layer was observed before it is presented to the AI agent for judgment.
[COMMAND_EXECUTION]: The skill makes extensive use of subprocesses to interact with system tools.
Evidence: scripts/pflib/verify.py and scripts/pflib/run.py invoke pdftotext, pdfinfo, and pdftoppm to process downloaded files. These calls use argument lists rather than shell strings, which mitigates standard shell injection risks.
Dispatcher: scripts/pf.py uses os.execvp to replace the current process with a uv run call for sibling scripts. This is restricted to scripts existing within the local scripts/ directory.
[DYNAMIC_EXECUTION]: The skill performs dynamic code execution within the user's browser environment.
Evidence: scripts/pflib/browser.py executes JavaScript snippets via opencli browser eval to handle complex web interactions and data transfers. These snippets are pre-defined templates, and URL inputs are safely escaped using json.dumps().
[EXTERNAL_DOWNLOADS]: The skill is primarily designed to download content from external sources.
Evidence: It fetches metadata from academic APIs (Crossref, OpenAlex, Europe PMC) and downloads PDF files from publisher domains. These interactions are paced and governed by a local pace.json file to comply with site policies.

paper-fetch