timesfm-forecasting
Pass
Audited by Gen Agent Trust Hub on Jun 26, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill downloads pre-trained model weights from official Hugging Face repositories (e.g., google/timesfm-2.5-200m-pytorch). This is a standard and necessary operation for using foundation models locally and utilizes a trusted hosting provider.\n- [COMMAND_EXECUTION]: The skill uses a dedicated preflight checker (scripts/check_system.py) that executes read-only system commands (sysctl, vm_stat) and utilizes ctypes to determine physical RAM and GPU availability. These are safety measures designed to protect the user's environment from hardware failure during model initialization.\n- [PROMPT_INJECTION]: The skill ingests untrusted data from user-provided CSV files for forecasting. While this represents a surface for indirect prompt injection, it is highly mitigated by the skill's architecture, which converts all input data to 32-bit floats (np.float32) before processing, effectively stripping any non-numeric instructional content.\n
- Ingestion points: CSV reading logic in scripts/forecast_csv.py and examples/anomaly-detection/detect_anomalies.py.\n
- Boundary markers: Data is treated as numeric arrays; no specific markdown or instruction delimiters are present for the data stream.\n
- Capability inventory: The skill utilizes subprocess calls for hardware detection and allows tool access for standard file operations.\n
- Sanitization: Strict numeric casting (astype(np.float32)) is applied to all ingested time-series data.
Audit Metadata