Google’s AI Push in India Signals a Fresh Start

When Google took the stage at I/O Connect India 2026, the buzz wasn’t just about flashy demos – it was about responsibility. The company announced a trio of AI‑centric initiatives designed to make the technology useful and safe for a country that’s sprinting toward digital transformation.

First up is Sec‑Gemini v3, a specialised AI security agent that Google is handing to a select group of trusted partners, including the e‑commerce giant Flipkart. Unlike a traditional firewall, Sec‑Gemini can sniff through massive streams of security data, spot anomalies, and even suggest remediation steps in real time. In a world where threats evolve faster than patches, that kind of speed feels almost necessary.

Google also pointed to a handful of internal projects that have already proven their worth. The “Big Sleep” research agent, built by DeepMind and Project Zero, has been used to flag live‑exploitation attempts – a first, according to Sandra Joyce, VP of Threat Intelligence. Meanwhile, the CodeMender agent can automatically draft patches for vulnerable code and push them straight to open‑source repositories, shrinking the window between discovery and fix.

But Google isn’t stopping at clever bots. The firm is open‑sourcing CAPSEM (Capabilities Security for Agents), a sandbox‑like runtime that isolates AI agents from the rest of a system. Think of it as a virtual safety cage: if an agent is coaxed into misbehaviour, the surrounding infrastructure stays untouched.

On the standards front, Google is lobbying for two W3C drafts that could change how agents interact with users and money. The Device‑Bound Session Credentials (DBSC) protocol would tie session tokens to a device’s hardware, rendering stolen cookies useless. The Agents‑to‑Payments (AP2) framework aims to let low‑value (under $100) AI‑initiated transactions happen securely, without a human having to click “approve” every single time.

While Google’s announcements were very India‑centric, the article also notes another milestone that could reshape the global AI scene. Mira Murati, the former OpenAI CTO who briefly led the company during its 2023 upheaval, has launched Thinking Machines Labs’ first model – Inkling. It’s a massive Mixture‑of‑Experts transformer with 975 billion total parameters (41 billion active) and a staggering 1 million‑token context window. A lighter sibling, Inkling‑Small, offers 12 billion active parameters for faster, cheaper inference.

Inkling’s open‑weight approach, coupled with its multimodal training on 45 trillion tokens of text, images, audio and video, signals a shift toward more transparent, fine‑tunable models. For Indian developers eager to tailor AI to local languages and use‑cases, that openness could be a game‑changer.

All of this points to a broader narrative: AI is no longer an experimental sandbox in Silicon Valley. It’s landing in classrooms, farms, hospitals, and government offices across India, and the tools to keep it secure are coming along for the ride. Whether the new agents live up to their promises will depend on how quickly developers, regulators, and users can adapt – but the ink is definitely drying on a fresh chapter.