hermes-workflows-creating
Pass
Audited by Gen Agent Trust Hub on Aug 1, 2026
Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill documents the use of a
bash(...)primitive for executing deterministic local commands within a workflow. This is an intended feature of the framework for tasks like checking environment status or recording evidence. - [PROMPT_INJECTION]: The authoring patterns described ingest external data which is then synthesized by LLM-based agent steps.
- Ingestion points: Workflow inputs (
MyWorkflowInput) and data retrieved from deterministic bash steps as shown in the example code in SKILL.md. - Boundary markers: The guide explicitly recommends using typed dataclasses for inputs and outputs to ensure data structure consistency and validation.
- Capability inventory: The framework includes capabilities for shell execution (
bash), LLM-based processing and synthesis (agent), and human-in-the-loop review requests (ask). - Sanitization: The skill mandates that authors implement human approval gates (
ask) for all external side effects and destructive actions (e.g., deployments, payments, PR merges), providing a critical manual verification step. - [SAFE]: Guidelines explicitly require recording comprehensive execution receipts, including stdout, exit codes, and hashes, to ensure transparency and auditability of all automated steps.
Audit Metadata