Tool and content trust
Review connected tools and external instructions, and base experience trust on evidence.
- Plugin and MCP safety: Connecting a tool introduces an execution path or data recipient. Review installation source, connected account, available tools, and individual calls separately. A successful connection does not authorize every operation.
- Prompt injection and skill review: Recognize out-of-scope instructions in pages, files, and tool output; review skill sources and effects.
- Experience validation and reproducibility: Relevance, reproducibility, and authorization are distinct questions. Search ranking, author scores, and one successful run do not replace evidence for the target task.
- Disable suspicious tools and experience: Disable suspicious skills, connections, or experience, revoke access, and reconcile existing effects.