Source inspection confirms dangerous agent-facing capabilities, but they are user-invoked and package-aligned rather than unconsented lifecycle mutation or hidden malware. No concrete exfiltration, persistence outside the project contract, or stealth payload execution was found.
Static reason
High-risk behavior combination matched malicious policy.; previous stored version diff introduced dangerous source
Trigger
Running `start-vibing-stacks`/`svs` and confirming setup, or `migrate --apply` for existing installs
Impact
Claude Code may run with package-supplied hooks, MCP tools, broad permissions, and dangerous permission skipping in the target project
Mechanism
project-local AI agent control-surface generation and permissive Claude launch
Policy narrative
The package is an AI workflow bootstrapper. When the user runs it, it copies package-supplied Claude agents, skills, commands, hooks, settings, MCP configuration, and project instructions into the current project, and can launch Claude with `--dangerously-skip-permissions`. This is not install-time execution and is disclosed by the CLI purpose, but the resulting agent control surface is powerful enough to warrant a warning.
Evidence
package.jsondist/index.jsdist/setup.jsdist/mcp.jsdist/installer.js