Static inspection shows guarded, user-invoked AI project bootstrap with high-risk Claude/MCP capabilities, not unconsented npm lifecycle control-surface hijacking. Warn because default permission-skip launch, project hooks, MCP registration, and Cursor skill installation create meaningful agent extension risk.
Static reason
High-risk behavior combination matched malicious policy.; previous stored version diff introduced dangerous source
Trigger
User runs the CLI, confirms setup, optionally accepts MCP defaults, or runs migrate --apply.
Impact
A configured Claude session may run project hooks and MCP tools with broad permissions; selected external tools may process project context or resolved environment values.
Mechanism
interactive Claude/Cursor project agent bootstrap with hooks, MCP registration, dependency installers, and permission-skipping launcher
Policy narrative
When invoked, the package interactively detects a project stack, copies packaged Claude agents/skills/hooks into the current project, writes Claude settings enabling project hooks and broad tool permissions, may register selected MCP servers, and then launches Claude Code with `--dangerously-skip-permissions`. It can also run system dependency installers. This is dangerous agent-facing capability, but it is not hidden install-time mutation and no secret theft or exfiltration path was found.
Evidence
package.jsondist/index.js