literature-review-tools
Pass
Audited by Gen Agent Trust Hub on Sep 4, 2026
Risk Level: SAFECOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The launcher script
scripts/litrun.pyfacilitates the installation of external Python packages from PyPI and clones code repositories from GitHub (e.g.,gpt-researcherandstorm) to launch web interfaces. While these targets are well-known research projects, the capability to download and execute remote code is a significant attack surface. - [COMMAND_EXECUTION]: The skill makes extensive use of
subprocess.runinscripts/litrun.pyto drive external command-line tools for PDF extraction, research synthesis, and environment management. It passes user-provided arguments through the launcher to these external processes. - [DYNAMIC_EXECUTION]: The
litrun.pyscript dynamically constructs and executes shell commands for tool management. Specifically, it can execute Python code snippets viapython -c(incmd_runfor libraries) and launches long-running UI servers usinguvicornorstreamlitbased on configurations inrecipes/recipes.json. - [INDIRECT_PROMPT_INJECTION]: The skill fetches content (titles, abstracts, and potentially full texts) from multiple external academic databases such as OpenAlex, PubMed, and arXiv.
- Ingestion points: Data enters the agent's context through
scripts/fetch_papers.py,scripts/fetch_arxiv.py,scripts/fetch_openalex.py,scripts/fetch_pubmed.py, andscripts/resolve_oa.py. - Boundary markers: The skill lacks explicit boundary markers or instructions to the LLM to ignore potentially malicious commands embedded within retrieved research data.
- Capability inventory: The skill possesses capabilities to write files, make network requests, and execute subprocesses through the
litrun.pylauncher. - Sanitization: The scripts perform minimal sanitization, such as stripping HTML tags from abstracts, but do not filter for adversarial prompt content.
- [EXTERNAL_DOWNLOADS]: The skill performs numerous downloads from external sources, including academic API endpoints (e.g.,
api.openalex.org,eutils.ncbi.nlm.nih.gov) and file downloads from repositories like arXiv and Europe PMC. It also clones repositories from GitHub for UI-based research tools.
Audit Metadata