ai-content-pipeline
Pass
Audited by Gen Agent Trust Hub on Apr 16, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill orchestrates multi-step workflows by executing
infshCLI commands in the shell. These templates use shell loops and variable interpolation to process data between different AI models. - [EXTERNAL_DOWNLOADS]: The documentation references external installation scripts and media assets hosted on vendor-controlled domains (
inference.sh) and GitHub repositories (github.com/inference-sh). These are documented as vendor resources for the core functionality. - [PROMPT_INJECTION]: The skill identifies patterns where untrusted external data (such as blog content or model-generated scripts) is interpolated directly into subsequent model prompts without boundary markers. This creates a surface for indirect prompt injection.
- Ingestion points: User-provided text inputs and intermediate JSON outputs (e.g.,
script.json,summary.json) inSKILL.md. - Boundary markers: Absent; data is placed directly inside JSON strings within shell commands.
- Capability inventory: Shell access for execution of AI applications via the
infshtool. - Sanitization: No sanitization or validation of the ingested content is defined in the instructions.
Audit Metadata