Consumer AI Agents: Privacy, Sharing & Permission Risks

Reports about Meta Muse share links describe roughly 6,800 links found through passive DNS, with conversation titles in URLs, alongside a separate account-control vulnerability; the findings and Meta’s response need verification. Shared-agent access-control proposals expose risks around derived files, tool outputs, and persistent memory. Claude Duet adds a signal of persistent, tool-connected assistants with approval gates, but reliability, permissions, and user value remain unproven.