Meta’s new agentic AI system, MetaMuse, is raising major alarm bells across the cybersecurity community by quietly provisioning unmetered, root-access cloud VMs for every user.

In this episode of the Security Boulevard Podcast, hosts Tom Hollingsworth and Fernando Montenegro sit down with Stephen Foskett (host of Utilizing AI) to dissect the surprising design choices behind MetaMuse. From running executable commands as root in a dedicated Ubuntu environment to aggressively seeking network connections like Tailscale and scraping personal data across Meta platforms, MetaMuse illustrates the dark side of "aggressively helpful" AI agents.

The panel breaks down Meta's "by design" philosophy, the failure of narrow threat modeling, the flaws of agentic guardrails like Sentinel, and the potential security fallout if fleet learning and vibe-coded AI package stores take off.

This and more on the Security Boulevard Podcast, part of the Futurum Podcast Network and Futurum Media.