Meta's Muse Spark 1.1 Turns Meta AI Agentic — and Publishes an 11.2% Multi-Turn Attack Success Rate That Defenses Only Cut to 6.9%

The Enterprise (corroborated by The Batch #363)·medium signalXBlueskyLinkedInCopy link

Meta shipped Muse Spark 1.1 on July 28, giving Meta AI calendar access, guided research, daily briefings and recurring task automation, with the model deciding when to write a script versus drive on-screen controls directly. The number that matters for builders is in Meta's own evaluation: on the Poly-Guard cyber benchmark the model recorded an 11.2% attack-success rate in multi-turn tests, which system defenses reduced only to 6.9%. Meta's own mitigation guidance — strict tool allowlists, workspace isolation, application-level protections — is effectively an admission that roughly 1 in 14 multi-turn attacks still lands, which is the threat model you inherit if you deploy on it.