Engineers at Meta scrambled to patch critical virtual machine security vulnerabilities in its new AI agent product, Muse, immediately prior to its public launch, according to internal documents obtained by 404 Media. The flaws included Linux kernel-based KVM escapes that could have allowed normal users to break out of isolated user environments and access sensitive internal Meta databases and core infrastructure. Internal sources report that security teams were forced to rush out rapid hotfixes to avoid delaying the launch.
Why it matters
Underscores severe infrastructure security risks inherent to hosting autonomous AI agents with execution privileges on corporate servers.
Warns engineering leaders against rushing agentic AI deployments without rigorous, multi-layered sandbox isolation boundaries.
Source: 404media.co



