earth2studio-create-prognostic
Pass
Audited by Gen Agent Trust Hub on Jul 22, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill provides instructions for the agent to manage a Python development environment and run tests using 'uv' and 'make'. Evidence: 'SKILL.md' contains commands like 'uv run pytest', 'uv sync', 'make format', and 'make lint'.\n- [EXTERNAL_DOWNLOADS]: The skill clones the Earth2Studio source code from NVIDIA's official GitHub repository during environment setup and accepts reference scripts from user-provided URLs. Evidence: 'evals/environment/Dockerfile' contains 'git clone https://github.com/NVIDIA/earth2studio.git', and 'SKILL.md' allows providing a URL via '$ARGUMENTS'.\n- [REMOTE_CODE_EXECUTION]: Implementation templates for model wrappers include the use of 'torch.load' with 'weights_only=False', which is a standard but sensitive ML pattern. Evidence: 'references/method-templates.py' and 'references/skeleton-template.py' provide templates for loading model checkpoints resolved from a 'Package' object (e.g., from HuggingFace or NGC).\n- [PROMPT_INJECTION]: The skill processes external scripts provided by the user, creating an indirect prompt injection surface. Evidence Chain: 1. Ingestion point: User-provided URL or path in '$ARGUMENTS' (SKILL.md Step 0). 2. Boundary markers: Absent. 3. Capability inventory: Writing files to the workspace and executing shell commands (uv, make). 4. Sanitization: Absent.
Audit Metadata