deepxiv

Warn

Audited by Gen Agent Trust Hub on Nov 10, 2026

Risk Level: MEDIUMDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [DYNAMIC_EXECUTION]: The skill implements a runtime resolution mechanism to locate and execute an adapter script (deepxiv_fetch.py). It searches across several paths, including $ARIS_REPO/tools/, local tools/, and ~/.codex/skills/deepxiv/. This dynamic path resolution and subsequent execution with python3 can be used to execute unauthorized code if an attacker can place a file in any of the checked locations.
  • [INDIRECT_PROMPT_INJECTION]: The skill retrieves academic paper content from external sources, creating a vulnerability to indirect prompt injection. * Ingestion points: Paper metadata, abstracts, and full-text sections retrieved via the deepxiv CLI or adapter script (SKILL.md, Step 3). * Boundary markers: The instructions do not define any delimiters or specific prompts to ensure the agent treats the paper content as untrusted data. * Capability inventory: The skill has the capability to execute shell commands (SKILL.md, Step 3) and write to local files (updating the Research Wiki in Step 6). * Sanitization: No sanitization, filtering, or validation of the external paper content is performed before it is presented to the agent or recorded in the project wiki.
  • [COMMAND_EXECUTION]: The workflow involves constructing shell commands that incorporate user-provided arguments such as search queries, paper IDs, and section names. The lack of explicit sanitization instructions for these variables poses a risk of command injection if the arguments are not correctly escaped by the underlying platform.
  • [EXTERNAL_DOWNLOADS]: The skill recommends installing the deepxiv-sdk package and python-dotenv. While identified as vendor-provided resources for the author, these represent external dependencies that are downloaded and installed at runtime.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 10, 2026, 09:37 AM