The package contains real dangerous agent-facing behavior, especially permission-bypassed Claude launch and broad project .claude writes, but activation is an explicit interactive CLI flow rather than install-time hijack. Per policy this warrants a warning rather than a publish block absent exfiltration, hidden persistence, or lifecycle-triggered foreign control-surface mutation.
Static reason
High-risk behavior combination matched malicious policy.; previous stored version diff introduced dangerous source
Trigger
Running `npx start-vibing-stacks` or `svs` and confirming setup or launch prompts.
Impact
Can grant a Claude Code session broad project tool permissions and install package-provided hooks/agents/skills into `.claude/`.
Mechanism
User-invoked Claude project configuration plus permission-bypassed Claude launch
Policy narrative
When the CLI is run, it detects the project, prompts for stack choices, copies Claude Code agents/skills/hooks into `.claude/`, writes permissive Claude settings, optionally configures MCP servers, and then launches `claude --dangerously-skip-permissions` unless disabled. This creates a broad AI-agent capability surface, but inspection did not find automatic npm lifecycle execution, stealth persistence, credential exfiltration, or unconsented install-time mutation.
Evidence
package.jsondist/index.js