senior-computer-vision

Warn

Audited by Gen Agent Trust Hub on Nov 12, 2026

Risk Level: MEDIUMDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [DYNAMIC_EXECUTION]: The inference_optimizer.py script uses torch.load() to process model checkpoints. This function is known to be unsafe as it utilizes the pickle module for deserialization, which can be exploited to execute arbitrary code if the model file is untrusted or malicious.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes external, user-provided datasets and model files, which constitutes a vulnerability surface for indirect prompt injection attacks.
  • Ingestion points: Untrusted data enters the agent's context through scripts/dataset_pipeline_builder.py (parsing JSON, XML, and TXT annotation files), scripts/inference_optimizer.py (loading and analyzing model files), and scripts/vision_model_trainer.py (reading dataset structures).
  • Boundary markers: The instructions do not define boundary markers or explicit delimiters to prevent the agent from interpreting instructions that might be embedded within large dataset annotation fields or model metadata.
  • Capability inventory: Across its scripts, the skill possesses the capability to write, copy, and symlink files to the filesystem and perform multi-scale model benchmarking and configuration generation.
  • Sanitization: The skill performs format-level validation (e.g., checking for valid JSON/XML structures) but lacks logic to sanitize or filter natural language instructions embedded within data fields.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 12, 2026, 02:12 AM