Google Cloud roundup: SentinelOne SecOps tie-up, Apigee AI guardrails and more

SentinelOne integration rebuilt for Google SecOps

Google Cloud and SentinelOne have reworked their security integration to handle threats at greater scale, according to the Google Cloud Blog. SentinelOne, a maker of endpoint detection and response software, is a Google Unified Security recommended partner.

The new setup connects SentinelOne directly to Google Security Operations using streamlined data schemas. Google says this gives SecOps detection and agentic security operations centre pipelines access to endpoint telemetry while lowering the total cost of running the integration.

Layered AI guardrails on Apigee

A new technical guide from Google describes how to stack three layers of protection for large language model traffic on Apigee, which acts as an AI gateway. The layers are deterministic static rules, Google Cloud Model Armour filters that work on meaning rather than fixed patterns, and business logic handled by NVIDIA NeMo Guardrails.

Google says the layered approach intercepted all benchmarked attacks mapped to the OWASP framework without blocking safe prompts. The layers run cheapest first, which the company says improves performance and cuts token costs. No single guardrail catches every threat, the guide notes.

Training and agent tooling on GKE and Cloud Run

Google has also published a blueprint for running distributed reinforcement learning post-training of Gemma 3 models on Google Kubernetes Engine. It combines GKE Custom ComputeClasses with DWS Flex-Start and Spot fallback, 8-NIC RoCE RDMA networking, NVIDIA NeMo RL and vLLM V1, with the stated aim of improving Gemma 3 mathematical reasoning accuracy.

A separate guide explains how to connect Gemini Enterprise privately to internal Model Context Protocol servers on Cloud Run. The setup uses Egress Agent Gateway, Private Service Connect, FQDN-scoped Private Cloud DNS for Shared VPCs, and OAuth 2.0 with IAM authentication. A reference implementation is available on GitHub.

Another Apigee article covers a stateful control plane for multi-agent workloads. Google claims features such as pre-warmed caching, trace compaction and dynamic model tiering between Gemini Flash and Frontier models can bring net FinOps cost savings of up to 80%. These are Google's own figures and have not been independently verified.

Anthropic runs its telemetry pipeline on Pub/Sub

A new case study describes how Anthropic, the company behind the Claude AI models, uses Google Cloud Pub/Sub for cross-cloud telemetry and billing ingestion. Anthropic needed a serverless messaging queue with no manual partition management and strict security, according to the post.

Using Pub/Sub with Private Service Connect, Anthropic built what Google calls a zero-ETL architecture handling more than 1 million events per second and 500 billion daily rows. The company reports an 88% reduction in consumer latency and sub-second end-to-end event delivery. Google has also published an on-demand webinar on architecting AI workloads with Pub/Sub.

New hardware, database tools and events

Google says A5X instances, accelerated by NVIDIA Vera Rubin NVL72, are now generally available as part of its AI Hypercomputer, which combines compute, storage, networking and software as one system.

Database administrators get a new browser-based Web UI for AlloyDB Omni, Google's self-managed PostgreSQL offering. The dashboard shows cluster health, hardware utilisation and active connections, and includes query diagnostics and performance tracking for the columnar engine across hybrid, multi-cloud and air-gapped environments.

Google is also running several Community TechTalks. A session on Model Armour for AI workload security is scheduled for 19 November 2026, and one on semantic model routing in Apigee using Antigravity took place on 12 November 2026. An earlier session on 29 October 2026 covered building governed LLM gateways in Apigee with Model Armour, Cloud Data Loss Prevention, rate limiting and response caching.