Meta launches personal agent Muse with safety deep-dive, but privacy VM is missing and data harvesting raises alarm
simplexmethod · x · 2026-09-16
Meta Superintelligence Labs launched Muse, a personal agent that runs inboxes, calendars and shells unattended, spawning subagents and building its own tools. The accompanying safety post details: model training for zero-shot tool calling, long context, prompt-injection awareness and multi-agent coordination; an assume-breach architecture with an isolated harness, no real credentials, and a non-overridable Sentinel gate on all external interactions; hardening via dogfooding, agentic red teaming and a newly opened bug bounty. Critics counter that Muse is generating vast actionable personal training data through its connectors — 'Onavo built in house' — while the Confidential VM promised to cryptographically block Meta's access isn't actually live yet, and most users would never enable it anyway.
Related event: Meta Unveils Personal Agent Muse, Sparking Privacy Concerns(2 posts)→
More from Companies & People
- Perplexity Hires Reverse Engineers in Berlin for First Time to Mine Its Search Index — bo_wangbo · 2026-09-17
- TypeSafe AI Launches Foundation Model Optimized for Automation, Drawns Investor Praise — hardimanjames · 2026-09-17
- Berkeley's Angjoo Kanazawa teaches machines to understand the physical world for the robotics era — jfiance · 2026-09-17
- Rumor: Mistral hacked, samples of unreleased tools leaked and full codebase up for sale — teortaxesTex · 2026-09-17
- Ex-DeepMind researcher co-founds Polyphron to train RL models that grow human tissue — vqctran · 2026-09-17
- MrBeast is hiring a Director of Insights — and the internet can't believe it's real — round · 2026-09-17