Meta launches personal agent Muse and details its safety architecture: isolated harness, Sentinel gateway, no real credentials
infoxiao · x · 2026-09-09
Meta Superintelligence Labs has launched Muse, a personal agent used internally since early 2026, and published a deep dive on its safety engineering.
What Muse does
- Knows the user, runs in the background, launches swarms of subagents
- Builds its own tools and edits itself; can run unattended over your inbox, calendar, and shell
- Model trained specifically for zero-shot CLI/skills tool calling, long context, long-trajectory instruction following with built-in prompt-injection awareness, and multi-agent coordination
Safety design: assume the agent is under attack
- Harness runs in an isolated cell with no access to real credentials
- Every external interaction passes through a "Sentinel" the agent cannot override
- Hardened via dogfooding, agentic red teaming, and a private bug bounty program with real adversarial findings
The team says Muse will still make mistakes, but far less frequently and with far less damage. Related components were also announced as open source.
Related event: Meta Launches Muse, an Always-On Personal AI Agent(68 posts)→
More from coding & agent
- Kolab Benchmark Uses Pro Screen Recordings to Show AI Agents Overestimate Collaboration Quality — minnesotanlp · 2026-10-27
- Dev argues Claude burns far more tokens than GPT models for the same coding work — haider1 · 2026-10-27
- Stop building agents that can't change models: stable contracts, external state, real evals — rseroter · 2026-10-27
- Cheap models can cost more in agent harnesses than frontier ones, researcher warns — vasuman · 2026-10-27
- Kanzo: A 5k-Line Rust Kanban Binary With MCP for Claude Code and Codex — allmudi · 2026-10-27
- 'A human approves everything' is a trap: approval fatigue and how to actually design safe agents — Efistoffeles · 2026-10-27