skill-iter-tune
Pass
Audited by Gen Agent Trust Hub on Sep 3, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external skill definitions provided by the user. These definitions are interpolated into prompts for other AI models (Claude and Agy) via maestro delegate. A maliciously crafted skill could contain hidden instructions designed to break out of the prompt context and influence the orchestrator's behavior.
- Evidence:
- phases/01-setup.md reads skill files from user-specified paths.
- templates/execute-prompt.md and templates/eval-prompt.md interpolate ${skillContent} directly into model prompts.
- [DYNAMIC_EXECUTION]: The skill dynamically generates shell commands for the maestro CLI and constructs instructions for a general-purpose sub-agent to modify local files.
- Evidence:
- phases/02-execute.md and phases/03-evaluate.md use an escapeForShell function to wrap dynamically built prompts into a maestro delegate command.
- phases/04-improve.md launches an Agent tool with a prompt generated from evaluation.suggestions to modify skill files.
- [COMMAND_EXECUTION]: The skill makes extensive use of the Bash tool to manage the file system and execute internal tooling.
- Evidence:
- phases/01-setup.md uses Bash to create a workspace and backup directory.
- phases/02-execute.md uses Bash to snapshot skill states and execute the maestro delegator.
Audit Metadata