Cloud Security, AI Agents, and Data Platforms Advance

Hyperscalers and platform vendors pushed a wave of security, AI, and data infrastructure updates, while new advisories highlighted active exploitation risks. The day saw improved visibility into data-plane logging, faster agent execution paths, expanded in-region AI model access, and broader database choices across clouds. Trends research quantified the acceleration in vulnerability discovery and exploitation, and an industry accord outlined governance for frontier AI systems.

Strengthening Visibility and Agent Governance

AWS introduced Event Coverage in CloudTrail, a console view that surfaces which services and resource types have Data Events enabled across accounts and organizations. By consolidating coverage into a single dashboard and enabling subscriptions directly from the interface, security and operations teams can identify and close logging gaps more quickly. The feature is available across all commercial Regions where CloudTrail operates, supporting efforts to detect object- and resource-level access, including potential exfiltration.

Google announced the Remote MCP Server in public preview, giving AI agents managed access to gcloud and bq commands via the Model Context Protocol. The service runs in a sandboxed environment without ambient credentials, enforces caller IAM and organization policies, and integrates with Cloud Audit Logging. Model Armor helps screen prompts and responses, while two exposed tools—run_gcloud_command and run_bq_command—cover management, observability, and BigQuery operations. It is free to use in preview aside from normal resource costs and requires enabling the Cloud CLI Execution API and appropriate IAM roles.

Cloudflare outlined new Agent Controls amid a traffic shift where AI agents generate a majority of requests. The company splits traffic into Search, Agent, and Training categories and uses cryptographic Web Bot Auth to verify agent identity. To address publisher revenue loss from agent traffic, Cloudflare is piloting Pay Per Use and a Monetization Gateway to enable per-request pricing and payment enforcement, alongside developer tools such as Markdown for Agents, WebMCP, and bandwidth visibility to reduce waste and support more efficient interactions.

To help agents execute tasks more responsively, Cloudflare rebuilt its Containers platform with a new scheduling policy and faster runtime. The Containers Update lets applications choose each sandbox’s image and compute instance at start time, dramatically cutting time-to-interactive (from roughly 4.05s to 648ms in independent tests). Durable Objects remain central, with each Container retaining a controller accessible via an API for lifecycle and rollout logic. A managed base image (cloudflare/debian-trixie) and native filesystem snapshots in public beta support rapid workspace preparation, sharing, and restoration.

Faster AI Experiences and Agent Tooling

AWS made OpenAI’s GPT-6 Astra available in a high-speed tier on Bedrock. GPT‑6 UltraFast targets latency-sensitive applications and, per OpenAI, can deliver up to 6x faster inference and around 300 tokens per second throughput. In parallel, Anthropic’s Claude Opus 5.5 and Sonnet 5 gained Claude In‑Region support for in-country processing on Amazon Bedrock in the UK (London) Region, addressing data residency and compliance requirements for regulated sectors.

AWS added bulk update and version check capabilities to its agent skills. The Agent Toolkit now supports aws agent-toolkit check-skill-updates to compare installed skills against registry versions and aws agent-toolkit update-skill --all to update all outdated skills in one operation. The commands build on existing CLI features around the MCP Server and skill management for coding agents and require AWS CLI v2.37.0 or later.

On the developer desktop, Microsoft brought native Linux container management to Windows Subsystem for Linux. WSL Containers introduces wslc.exe (with container.exe as an alias) and fills previous gaps such as restart, file copy, health checks, network connect/disconnect, real-time events, mounts, and configurable storage. An API enables Windows apps to launch and interact with Linux containers; enterprise controls integrate with Microsoft Defender for Endpoint and Intune, and Docker Compose–style compatibility is planned.

Data Platforms, Analytics, and Compute Upgrades

Google announced the general availability of Spanner Omni, a self-managed, deploy-anywhere version of its distributed SQL database with converged SQL, graph, key-value, vector search, full-text, and columnar analytics. The GA release adds enterprise security (TLS, authN/Z), audit logging, high-performance backup/restore, and dedicated stateless worker nodes, while noting that SLAs are not provided for customer-managed deployments and certain managed-cloud integrations are not included.

Amazon expanded analytical reach for operational databases: Aurora PostgreSQL can now query Apache Iceberg and Parquet data in-place from Amazon S3, Amazon S3 Tables, and AWS Glue Data Catalog, using an embedded DuckDB engine. Customers can also federate to IRC-compatible catalogs via Glue and optionally materialize external data into native tables for lower-latency queries. The feature is generally available in all commercial and GovCloud (US) Regions at no additional charge for supported Aurora PostgreSQL versions.

For vector search applications, Amazon added metadata pre-filtering to S3 Vectors. S3 Vectors can now evaluate constraints before similarity search to return up to 5x more matching vectors for selective filters. A $startsWith operator simplifies prefix matches on structured values. New indexes enable pre-filtering by default, existing indexes can be migrated with UpdateIndexMode, and deployment is rolling out across all commercial and China Regions at no additional cost.

Amazon Managed Grafana advanced dashboard lifecycle and query flexibility. With Managed Grafana now supporting Grafana 13.2 workspaces, administrators gain Git Sync for dashboards-as-code workflows and dynamic dashboards that adapt to data and variables. The Amazon CloudWatch data source plugin adds PromQL for metrics ingested via the OTLP endpoint, complementing existing query types.

For bursty workloads, AWS enhanced serverless database elasticity. Aurora Serverless can now add up to 16 ACUs within a second and continue scaling up to 256 ACUs as needed, benefiting agentic AI and other spiky patterns while retaining scale-to-zero behavior. The capability is enabled by default for platform versions 3 and 4; older clusters can upgrade to version 4.

AWS broadened instance options for managed databases by adding AMD-based classes. R8a Instances (Aurora and RDS) and M8a Instances (RDS) use 5th generation AMD EPYC processors with one vCPU per physical core for predictable per-core performance, offer up to 75 Gbps network and 60 Gbps EBS bandwidth, and run on sixth-generation Nitro Cards. Regional availability spans multiple geographies; supported engine versions and pricing are documented by AWS.

For graphics-heavy desktop virtualization, AWS added a new GPU-backed instance family to WorkSpaces Core Managed Instances. WorkSpaces G7 uses NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs and Intel Xeon 6 processors, targeting CAD/CAM, 3D rendering, visualization, video editing, and AI-assisted design. AWS offers six sizes (1–8 GPUs), 8–192 vCPUs, and 32–768 GB system memory, with 32 GB GDDR7 per GPU and 2.67x higher memory bandwidth than the prior generation, initially in selected US and EU Regions.

Active Exploits, Exposure Trends, and Governance

Google’s Threat Intelligence Group analyzed CVE disclosures and exploitation from January 2025 to August 2026. The GTIG Report found monthly CVE counts roughly doubled in 2026 and observed exploited vulnerabilities rising from an average of 10.5 per month in 2025 to 18 per month in 2026, with growth largely driven by rapid weaponization of n-days. AI-assisted discovery tended to produce fewer low-severity issues and a higher share of RCE-capable findings; despite volume increases, exploited CVEs remained a small fraction of disclosures. The report recommends moving toward intelligence-driven triage and automated remediation.

Cisco warned that a zero-day in Catalyst SD-WAN Manager is being exploited. Cisco SD‑WAN CVE-2026-76504 allows remote, unauthenticated attackers to bypass authentication due to improper URI encoding handling in session login API requests, granting admin-level access. Cisco released fixed builds across multiple trains and advised restricting Manager exposure until upgrades are applied; no workaround exists.

CISA issued an alert for a critical MikroTik RouterOS flaw (CVE-2026-84411). The MikroTik RCE vulnerability is an integer underflow in HTTP request body handling that permits pre-authentication root code execution or denial of service via a single crafted request. CISA urged updates to vendor-recommended versions, network isolation of control systems, and secure remote access practices; no active exploitation evidence was cited at advisory time.

An industry agreement set out voluntary safety controls for frontier AI. The White House’s Safety Accord on Super Intelligence—signed by President Trump and six AI companies—commits to layered controls: internal monitoring and verification, independent external assessment, and board-level oversight. The accord highlights cybersecurity, biosecurity, and chemical risks, is nonbinding legally, and envisions continued work toward shared standards that could inform future regulation.