workflow-from-eval
Pass
Audited by Gen Agent Trust Hub on Oct 16, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The skill uses system utilities to process media and manage network communications.
- It utilizes
ffmpegandffprobefor extracting frames from video files, analyzing metadata, and generating visual review sheets (contact sheets). - It uses
curlto perform authenticated HTTP PUT operations for uploading data to cloud storage via signed URLs. - [EXTERNAL_DOWNLOADS]: The skill is designed to fetch and process user-provided media from external sources.
- It downloads images and videos from user-specified URLs into a local working directory (
frames/) for evaluation purposes. - [DYNAMIC_EXECUTION]: The skill generates and executes temporary scripts and logic blocks to facilitate workflow creation.
- It creates a helper shell script (
put.sh) to manage binary file uploads. - It dynamically generates computer vision workflow manifests and logic (via
expressionblocks) based on model outputs and identified business goals. - [DATA_EXFILTRATION]: The skill transmits data to vendor-controlled endpoints (Roboflow) as part of its core functionality.
- It manages the upload of processed assets to Roboflow's storage infrastructure using signed URLs.
- The instructions explicitly direct the agent to avoid logging or storing sensitive tokens, API keys, or signed URLs in session reports or persistent files, demonstrating a secure-by-design approach.
Audit Metadata