Meta opens public bug bounty for Muse and details its agent safety architecture

shuchaobi · x · 2026-09-09

Meta officially announced that the private bug bounty for Muse, its personal agent, has run since early development and is now public, with published payout guidelines paying by demonstrated impact. Alongside comes a 20-minute read, "How We Built Safety Into Muse," by Tarek Sheasha, VP at Meta Superintelligence Labs.

Muse launched today as a personal agent that actually does things: runs in the background, launches swarms of subagents, builds its own tools, and edits itself — the first time Meta handed inboxes, calendars, and a shell to unattended software.

The safety design assumes the agent may be under attack at all times:

The model is trained for zero-shot CLI tool calling, long context, long-trajectory instruction following with inherent prompt-injection awareness, and multi-agent coordination
The harness runs in an isolated cell and never sees real credentials
Every external interaction passes through a Sentinel the agent cannot override
Hardened via extensive dogfooding, agentic red teaming, and real adversarial findings from the private bounty

Meta acknowledges Muse will still make mistakes, but expects them to be far less frequent and less damaging.

Original post →

Meta opens public bug bounty for Muse and details its agent safety architecture

More from coding & agent