Google DeepMind / Google · Mountain View, CA, USA · founded 1998 · Gemini, Vertex AI, AX (agent orchestrator), Chrome AI
Google's AI-security notes are mostly about the security posture of the platforms it ships — the default configuration of its open-sourced agent orchestrator, and indirect prompt-injection research against Gemini-class assistants.
Reported incidents (2)
Google open-sourced AX, a Kubernetes-native agent orchestrator designed around sandboxing and an explicit network allowlist. An independent review (Wavect) reported that the documented example Gateway allowed every host on port 443 rather than a real allowlist, and that the budget/approval policy field was reserved and effectively disabled — a gap between the design intent and the shipped defaults. PlayCISO compared it against rival orchestrators.
Researchers have repeatedly demonstrated indirect prompt injection against assistant products that read untrusted content (email, documents, web pages), a class that applies to Gemini-integrated surfaces as to all LLM assistants. These are research demonstrations of a systemic weakness, not a single breach.