comm-lit-review

Pass

Audited by Gen Agent Trust Hub on Nov 10, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process large amounts of untrusted data from external sources, making it a target for indirect prompt injection.
  • Ingestion points: WebFetch tool fetches content from scholarly sites (IEEE, ScienceDirect, ACM) and broader web search results; the skill also reads user-controlled local PDFs and notes from Zotero/Obsidian.
  • Boundary markers: The skill lacks explicit boundary markers or instructions to the agent to ignore potentially malicious directions embedded within research paper abstracts or full-text fetches.
  • Capability inventory: The skill possesses extensive capabilities including Bash(*), Write, and networking tools, which could be abused if the agent interprets instructions within a retrieved document as authoritative.
  • Sanitization: There is no evidence of sanitization or filtering of the external data before it is processed into the final literature table.
  • [COMMAND_EXECUTION]: The skill requests and utilizes Bash(*) permissions to facilitate local file scanning (e.g., searching for PDFs in papers/ or literature/). While the instructions focus on benign file operations using locate, glob, and grep, the granting of unrestricted shell access provides a platform for arbitrary command execution if the agent is manipulated.
  • [DATA_EXFILTRATION]: The skill allows the library search path (PAPER_LIBRARY) to be customized via the AGENTS.md file. A malicious configuration targeting sensitive system directories, combined with the skill's instruction to 'extract relevance' and 'main results', could lead to the exposure of sensitive local data in the final research summary output.
  • [PROMPT_INJECTION]: Untrusted input from $ARGUMENTS is directly interpolated into the research topic heading and parsed for source selection directives. This provides a direct surface for users to attempt to override skill instructions or influence agent behavior through injection strings in the query.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 10, 2026, 09:36 AM