benchmark-pyrefly

Pass

Audited by Gen Agent Trust Hub on Sep 19, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • Command Execution: The skill facilitates the execution of build and benchmarking tools, including buck2 and cargo, which are used to compile and run performance tests. It also references a local maintenance script (update_revision.sh) for managing project dependencies.
  • External Data Ingestion (Indirect Prompt Injection Surface): The skill processes a large external codebase (PyTorch) for performance testing, which represents a potential surface for indirect instructions.
  • Ingestion points: The skill accesses a multi-gigabyte checkout of the PyTorch repository containing approximately 15,000 Python files across all cores during benchmarks.
  • Boundary markers: No explicit boundary markers or instructions to ignore embedded content within the external source code are specified in the provided files.
  • Capability inventory: The skill utilizes shell-based build systems (buck2, cargo) to compile and execute code, and provides a path for running local shell scripts for maintenance.
  • Sanitization: No specific sanitization, filtering, or validation of the external source code content is mentioned prior to the benchmarking operations.
  • External Resource Acquisition: The skill describes fetching external source code from public repositories and internal storage systems to support the benchmarking process. These operations are part of the standard development and testing workflow for the project.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 19, 2026, 11:29 AM