senior-computer-vision

Pass

Audited by Gen Agent Trust Hub on Nov 5, 2026

Risk Level: SAFEDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [DYNAMIC_EXECUTION]: The script scripts/inference_optimizer.py uses the torch.load function to analyze model checkpoints. This function can lead to arbitrary code execution if used on malicious pickle-based files, as the script does not implement restricted loading (e.g., weights_only=True). While this is a standard method in model engineering, it remains a security risk for untrusted files.
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests external data in the form of dataset annotations, which constitutes an injection attack surface. 1. Ingestion points: The tools scripts/dataset_pipeline_builder.py and scripts/vision_model_trainer.py read local JSON, XML, and TXT files provided by the user. 2. Boundary markers: There are no explicit markers or safety instructions to prevent the model from interpreting instructions embedded in annotation fields. 3. Capability inventory: The skill allows for file system operations and suggests shell command execution for model training. 4. Sanitization: The scripts validate data structure during parsing but do not sanitize the content of data strings against adversarial instructions.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 5, 2026, 08:53 PM