Agent Launches Expose VM‑Escape Risk

Shipping agentic AI that runs user‑owned code or tasks on provider infrastructure creates a new class of attack: VM (virtual machine) escape from user agent instances into host infrastructure. Firms rushing launches can produce temporary, incomplete mitigations that increase the chance of large data breaches or cross‑tenant compromise. — This reframes many product‑security incidents from isolated bugs into a structural governance issue: agent deployments inherently expand an attack surface that regulators, boards, and security teams must treat as a public‑safety risk.