Google Research outlines agentic AI privacy and security roadmap

Google’s new agentic privacy and security agenda arrives just as we are learning how brittle today’s agent stacks really are. The CAPS workshop report argues that useful agents must reason about social context, not just tokens, and proposes contextual policy engines that sit above models to decide whether a given action or data flow is appropriate. In parallel, independent work highlighted by Ars Technica shows how current protocols like MCP can be abused to bounce malicious prompts between agents, bypassing LLM‑level guardrails.

This combination is a warning to every lab racing to build autonomous research assistants and enterprise copilots. As agents gain the authority to move money, modify code and touch sensitive data stores, security failures stop looking like chat misfires and start looking like classic distributed‑systems breaches. Google is effectively saying that frontier capability work now has to be coupled with an equally serious program in contextual security, multi‑agent sandboxes and zero‑trust‑style designs for agent networks.

From a race‑to‑AGI standpoint, strong agent security is both accelerator and brake. In the near term, engineering robust supervisor layers and benchmarks will slow down productization of the wildest agent architectures. Over a longer horizon, labs that solve these problems will be able to deploy much more capable systems into high‑stakes domains without losing regulatory or customer trust, which ultimately supports faster and broader real‑world use of proto‑AGI agents.