skills/nvidia/warp/warp-eval/Gen Agent Trust Hub

warp-eval

Pass

Audited by Gen Agent Trust Hub on Aug 17, 2026

Risk Level: SAFECOMMAND_EXECUTION
Full Analysis
  • [DYNAMIC_EXECUTION]: The script scripts/measure.py uses subprocess.run to execute Python benchmarking drivers in isolated processes. This is a standard practice for performance measurement. The tool implements safety checks such as _driver_path (line 396) which ensures the driver script is within the local workspace using resolved.is_relative_to(workspace). It also utilizes _extra_environment (line 416) to filter out potentially dangerous environment variables like LD_PRELOAD, DYLD_LIBRARY_PATH, and PYTHONPATH, preventing library injection attacks.- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted code and documentation from a target repository for performance analysis, creating a potential injection surface. Evidence chain: (1) Ingestion points: The skill manifest SKILL.md defines the input as the 'target repository' and 'representative datasets'. (2) Boundary markers: The skill enforces an 'Authorization checkpoint' (references/authorization-checkpoint.md) and 'Rejection Gates' (references/rejection-gates.md) to ensure work only proceeds under explicit user approval and within safe bounds. (3) Capability inventory: The skill can perform local file reads and command execution via subprocess.run. (4) Sanitization: The measure.py utility implements strict path validation and environment sanitization for all executed scripts.- [METADATA_POISONING]: Analysis of the skill's metadata fields in SKILL.md and skill-card.md confirms that all claims regarding functionality, authorship (NVIDIA Corporation), and licensing (Apache-2.0) are consistent with the provided codebase. No deceptive or misleading metadata was identified.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 17, 2026, 05:53 AM
Security Audit — agent-trust-hub — warp-eval