build-iterated-agentic-loop

Warn

Audited by Gen Agent Trust Hub on Jun 23, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTIONCREDENTIALS_UNSAFEDATA_EXFILTRATION
Full Analysis
  • [COMMAND_EXECUTION]: The skill generates GitHub Actions workflows that execute shell commands and high-privilege CLI tools.
  • The helper script references/agent-iteration.ts uses shell interpolation with the GitHub CLI (gh) to fetch pull request data.
  • Generated workflows utilize flags such as --permission-mode bypassPermissions, --sandbox danger-full-access, and --dangerously-skip-permissions to allow agents to operate autonomously in CI runners.
  • The workflow template uses user-contributed comment bodies as command-line arguments for the iteration script, which represents a potential command injection risk, although this is restricted to repository members.
  • [EXTERNAL_DOWNLOADS]: The skill templates perform global installation of various coding agent packages from external registries.
  • Downloads include @anthropic-ai/claude-code, @openai/codex, opencode-ai, and @humanlayer/cli from NPM and Bun registries.
  • Standard GitHub Actions are used for environment setup, including actions/checkout and oven-sh/setup-bun.
  • [PROMPT_INJECTION]: An indirect prompt injection surface exists in the iteration workflow where untrusted data is processed.
  • Ingestion points: The references/agent-iteration.ts script fetches issue comments and review comments from the GitHub API.
  • Boundary markers: The generated prompt uses Markdown headers (e.g., '# PR Issue Comments') to delineate the external data.
  • Capability inventory: The configured coding agents are granted write permissions to repository contents and pull requests, and can execute shell commands within the CI runner.
  • Sanitization: There is no evidence of sanitization or filtering of the external comments before they are interpolated into the agent's prompt.
  • [CREDENTIALS_UNSAFE]: The skill relies on sensitive API keys and tokens for operation.
  • Workflows are configured to use ANTHROPIC_API_KEY, OPENAI_API_KEY, and GITHUB_TOKEN.
  • These secrets are passed into the CI environment as environment variables for use by the coding agents.
  • [DATA_EXFILTRATION]: The skill's primary function involves reading repository content and sending it to external AI service providers (Anthropic, OpenAI) for processing and code generation.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Jun 23, 2026, 06:08 PM
Security Audit — agent-trust-hub — build-iterated-agentic-loop