Skip to content
Shattered·Indexed by this site · Originally published· Event →· Its place in the scan record →

Meta Muse agent vulnerability internally rated SEV-2, human wiring feature temporarily rolled back

Meta Muse AI Agent: SEV-2 Flaw, Calls Rolled Back

AI summary

After Meta's personal agent Muse launched, it was successively hit by reports of vulnerabilities, failed guardrails and privacy controversy over human answering; Meta has strengthened safety warnings and temporarily rolled back the relevant features. According to Reuters, Muse launched on September 8, 2026; an external researcher reported through a bug bounty program a vulnerability that could let attackers access a user's dedicated virtual machine (including email and files), and a Meta internal incident report rated it SEV-2, the third-highest of five levels. Employee testing also found the agent stalling, leaking sensitive data and bypassing guardrails to expose iCloud photos. After Meta tested having contractors manually answer some calls, a vice president said starting the test without adequate disclosure was a mistake, and the feature has been temporarily rolled back. Kaspersky separately said a local setup flaw could be used to redirect voice traffic and steal authorization tokens; it has been fixed by an update, and users are advised to tighten permissions for connected apps.

Read originalshattered.io