loop-architect
Pass
Audited by Gen Agent Trust Hub on Oct 15, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [COMMAND_EXECUTION]: The skill uses subprocess calls to interact with AI model CLIs and execute verification scripts.
- Evidence: The
looper.pyscript andrun-loop.pytemplate utilizesubprocess.runto call tools likekiro-cli,claude, andollama, as well as user-defined programmatic checks. - Security Control: The skill strictly mandates the use of argument arrays (
argv) rather than shell strings for all command executions to prevent shell injection vulnerabilities. - [INDIRECT_PROMPT_INJECTION]: The skill has a defined surface for indirect prompt injection as it processes data from external files and model outputs to drive workflow logic.
- Ingestion points:
process-notes.md(user input) and responses from AI council members (reviewers and judges). - Boundary markers: The
run-loop.pyscript requires judges to return structured JSON within markdown fences to isolate the verdict from conversational text. - Capability inventory: The runner can execute arbitrary commands defined as 'programmatic checks' in the
loop.yamlconfiguration and call external model CLIs. - Sanitization: A sophisticated 'scrub layer' is implemented in
run-loop.pywhich scans prompts for content from sensitive files (matching globs like.envor*.key) and redacts them before sending data to models. - [EXTERNAL_DOWNLOADS]: The skill refers to official installation paths for CLI tools from well-known technology providers.
- Evidence:
looper.pycontains probe and installation metadata for services such as Google Gemini, Anthropic Claude, and Ollama. - Context: These references are informational and directed toward established official repositories and domains, posing no inherent security risk.
Audit Metadata