skills/akillness/jeo-skills/pixelrag/Gen Agent Trust Hub

pixelrag

Pass

Audited by Gen Agent Trust Hub on Oct 30, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill executes the pixelrag command-line interface to perform visual indexing and retrieval tasks. It also includes a local helper script, scripts/preflight.py, which is used to verify the integrity of the upstream repository using SHA-256 hashes.
  • [EXTERNAL_DOWNLOADS]: The skill facilitates the installation of the pixelrag Python package and its dependencies from standard package registries. It also references the download of large model weights and prebuilt search indexes (e.g., a 217 GB FAISS index) from the official project repository on GitHub.
  • [INDIRECT_PROMPT_INJECTION]: As a retrieval-augmented generation (RAG) tool, the skill processes untrusted external data such as web pages and PDF documents. This content is rendered into visual tiles for indexing, which could theoretically contain malicious instructions targeting a downstream vision-language model (VLM).
  • Ingestion points: External URLs and document paths provided to the pixelshot and pixelrag index commands.
  • Boundary markers: The skill documentation does not specify boundary markers for visual content, though it treats indexing as a multi-stage pipeline.
  • Capability inventory: The skill uses subprocess calls to the pixelrag CLI, performs network operations to fetch source data, and writes index artifacts to the local filesystem.
  • Sanitization: The skill relies on visual rendering (screenshots) as a primary data processing step, which provides a degree of isolation from traditional text-based prompt injection, though visual content remains a surface.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 30, 2026, 09:02 AM