AWS set the tone today with a wave of platform and security updates. Its generative AI service, Bedrock, added 18 fully managed open‑weight models to broaden vendor choice without changing application code. The company also previewed Security Agent to embed continuous, policy‑driven validation and context‑aware testing into development and deployment. The mix underscores a push toward flexibility and proactive controls, even as researchers detail a long‑running browser‑extensions campaign impacting millions of users.
Security and governance turn up
For cloud security operations, AWS made its unified cloud security service, Security Hub, generally available with near real‑time risk analytics and attack‑path visualizations that correlate findings from GuardDuty, Inspector, and CSPM. In parallel, Bedrock AgentCore added preview features for policy enforcement and automated evaluations, allowing teams to intercept tool calls in real time and gate agent quality with built‑in and custom metrics surfaced through CloudWatch. Together, these updates aim to prioritize risks and harden agent workflows without custom glue code.
At the application layer, the previewed AWS security agent from the lead moves policy checks earlier, while the new DevOps Agent autonomously triages incidents and recommends operational improvements across distributed environments. Organizations can evaluate both previews in US East (N. Virginia) and align access, logging, and governance with existing controls. The result is a tighter feedback loop that reduces manual correlation and periodic testing bottlenecks.
Models and platform choice expand
AWS broadened model choice with the largest single expansion of managed open‑weight options in its catalog, creating room to balance cost, latency, and accuracy across workloads without refactoring. It also introduced Nova 2 on the service for step‑by‑step reasoning, tool use, and a million‑token context, plus thinking‑intensity controls to tune speed and cost. For teams building deeper customization, AWS made Nova Forge generally available to develop models from Nova checkpoints using reinforcement fine tuning and a responsible‑AI toolkit, blending proprietary and curated datasets while preserving core capabilities. These moves consolidate options for experimentation and production without sacrificing governance.
Scaling AI operations and data
On the data plane, AWS announced general availability of S3 Vectors, introducing vector buckets that scale to billions of embeddings with default encryption and optional KMS keys for multi‑tenant and regulatory needs. Complementing storage, OpenSearch added GPU‑accelerated index building and serverless auto‑optimize jobs to cut build time and cost and remove weeks of manual tuning. The combination targets faster iteration for retrieval‑augmented generation, semantic search, and agent memory at production scale.
For deployment strategies in sovereign or regulated settings, AWS introduced AI Factories—managed, high‑performance AI infrastructure delivered into customer data centers with physical and operational separation. To accelerate training, it also unveiled Trn3 UltraServers powered by fourth‑generation Trainium3, offering substantial compute, memory bandwidth, and interconnect gains for frontier‑scale models and expert‑parallel workloads. These offerings aim to reduce time‑to‑capacity and improve price‑performance while aligning with sovereignty and lifecycle requirements.
Operational building blocks evolved as well. AWS introduced Lambda durable functions for long‑running, multi‑step applications and AI workflows with platform‑managed state, checkpointing, and failure recovery. And MLflow went serverless in SageMaker AI, removing the need to provision and maintain tracking servers and enabling cross‑account collaboration via RAM. Together, these updates trim operational overhead and make it easier to compose reliable, cost‑efficient pipelines.
Confirmed extensions campaign
Researchers documented a multi‑year browser‑extensions operation that began with utility and wallpaper add‑ons and later pivoted to surveillance and remote code execution at scale. According to Infosecurity, one cluster’s Chrome extensions used a backdoor to fetch arbitrary JavaScript and exfiltrate encrypted browsing histories and fingerprints, affecting about 300,000 users; on Microsoft Edge, spyware‑laden add‑ons—including WeTab—reached more than four million installs, collecting visited URLs, search terms, mouse clicks, cookies, keystrokes, and identifiers. Many were removed after discovery, but the campaign’s longevity highlights marketplace review gaps. Recommended mitigations include auditing installed extensions, removing unused add‑ons, preferring developers with transparent update histories, and monitoring permission changes for unusual behavior.