Critical RCE vulnerability hits Google Cloud ADK

The biggest security threats to AI agents often lie in traditional software engineering oversights rather than rogue model reasoning or prompt injection. Attackers do not need to jailbreak or trick the LLM because the flaw executes directly through framework tooling and session replay logic. The exploit relies on pytest residing in the production environment, reinforcing why test harnesses must be purged from production container images. Receiving a CVSS 10.0 rating, the vulnerability exposes standard container runtimes like Cloud Run and GKE to unauthenticated RCE, requiring developers to update ADK to a patched release immediately and strip test packages from production builds.