The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill features a Data Processing Pipeline and a Conditional Workflow that ingest untrusted external data.\n
Ingestion points: $(cat $file) in data_processing.sh and INPUT_TEXT="$1" in conditional_workflow.sh.\n
Boundary markers: Absent; data is interpolated directly into the JSON input for the AI model without delimiters or instructions to ignore embedded commands.\n
Capability inventory: The skill executes subprocess calls to the belt CLI to run various AI models and services.\n
Sanitization: Absent; there is no escaping or validation of the ingested content before it is passed to the AI prompt.\n- [REMOTE_CODE_EXECUTION]: The Quick Start section directs users to a remote installation script hosted on GitHub (inference-sh/skills) to install the required CLI tool, which involves executing remote content.\n- [PERSISTENCE_MECHANISMS]: The skill provides explicit instructions and templates for setting up persistence via crontab entries to schedule recurring AI automation tasks.\n- [DYNAMIC_EXECUTION]: The shell templates demonstrate unsafe variable interpolation (e.g., "prompt": "$prompt...") which can lead to argument injection if inputs contain shell metacharacters. Furthermore, the run_with_alert utility function executes its arguments directly using $@, providing a pattern for arbitrary command execution.

ai-automation-workflows