Cloud Resilience, Edge Runtime Expansions, and Key Security Fixes

Cloud platforms rolled out capabilities aimed at scaling AI workloads, hardening operational resilience, and simplifying governance. At the same time, researchers and vendors flagged active exploitation and patched flaws across identity, software supply chains, and specialized lab systems. The updates below focus on practical impact for engineering and security teams, with emphasis on validated sources and actionable details.

AI Scale and Customization on AWS

AWS Bedrock added support for 1 million token context windows for OpenAI GPT-5.6 Sol, Terra, and Luna. The long-context capability is designed to keep full inputs—entire repositories, lengthy legal or regulatory documents, or complete agent histories—in a single request, reducing chunking and preserving coherence. To manage repeated analysis costs, Bedrock introduced prompt caching with explicit breakpoints, billing repeated context at a 90% discount. Pricing is stated to match OpenAI first‑party rates and counts toward existing AWS commitments. Availability varies by model: Sol in US East (N. Virginia) and US East (Ohio); Terra and Luna in US East (N. Virginia), US East (Ohio), and US West (Oregon). The models are accessible via the Bedrock Console or the Responses API on the bedrock‑mantle endpoint.

SageMaker expanded AI serverless model customization to support full fine‑tuning for more than 25 open‑source models, moving beyond parameter‑efficient approaches such as LoRA. Supported families include gpt‑oss, Gemma, Llama, Nemotron, and Qwen, enabling updates to all parameters for domain adaptation, terminology alignment, specialized reasoning, and complex output formats. The serverless workflow handles infrastructure provisioning and training orchestration so teams pay only for consumed resources. Regional availability spans US East (N. Virginia), US West (Oregon), Asia Pacific (Tokyo), and Europe (Ireland), with initiation via SageMaker Studio or the Python SDK.

Cloudflare Extends Edge Runtime for Agents and gRPC

Cloudflare introduced private beta features that bring inbound TCP and native gRPC support to Workers. A new connect() handler exposes raw sockets and allows passing sockets between Workers, Durable Objects, and Containers, enabling full‑duplex TCP protocols. A Spectrum application type routes incoming TCP traffic to a Worker, positioning Cloudflare as an ingress proxy for non‑HTTP services. The platform also translates between gRPC and gRPC‑web for browser and Worker compatibility, leveraging 330+ edge locations to reduce latency for workloads such as real‑time voice AI and streaming inference.

Cloudflare also unveiled an early preview of @cloudflare/computer, an open‑source agent runtime that gives each agent a virtual computer: a durable filesystem and multiple execution backends. The Workspace, backed by SQLite, can be populated from sources like git repositories or storage buckets and accessed via an API or a node:fs‑compatible wrapper. An AI SDK‑compatible toolkit exposes read, write, edit, ls, and exec tools; exec selects an isolate‑based runtime for speed or a full container when native binaries or a Linux environment are required. The architecture separates the agent harness from sandboxed execution, emphasizes auditing and gating of file operations, and aims to keep container usage to a minority of tasks while scaling agents across isolates and Durable Objects.

AWS Resilience, Governance, and Throughput Updates

Resilience Hub introduced recommended resilience tests to automate validation against predefined failure scenarios. The capability generates pre‑configured tests tailored to a service’s architecture and policy, uses AWS Fault Injection Service (FIS) to inject faults (e.g., Availability Zone or Regional impairments and dependency failures), evaluates outcomes against alarms and recovery objectives, and returns pass/fail results with detailed reports. It is available across a broad set of regions and supports continuous assurance without extensive manual test design. In parallel, Service Quotas now exposes the maximum account quota and utilization for AWS Organizations. Administrators can view and monitor usage from the management account via the console or GetServiceQuota API—initially in US East (N. Virginia)—allowing proactive requests for increases and avoiding account‑creation disruptions.

AWS Config added support for 15 additional resource types across services including Amazon Bedrock, Amazon OpenSearch Serverless, SageMaker, AppSync, Connect, and Glue. The expansion, effective in all regions where the resources exist, brings more assets into inventory and compliance evaluation, with support in Config rules and aggregators for broad discovery, audit, and remediation workflows. For application‑layer protections, AWS WAF now supports two Miggo Security managed rule groups—High Emerging Application Threats and AI/ML Application Protection—delivering continuously updated signatures for vulnerabilities that are actively exploited, have public proof‑of‑concepts, or appear in CISA’s KEV catalog. Subscriptions are added to web ACLs directly from the console, with versioning support and pricing via AWS Marketplace.

AWS Lambda increased the limit for Provisioned Mode event pollers on Amazon SQS event source mappings from 2,000 to 10,000, enabling up to 100,000 concurrent invocations per mapping. The change targets latency‑ and throughput‑sensitive workloads—such as real‑time order processing, financial transaction pipelines, IoT ingestion, and large‑scale fan‑out—reducing the need to shard across multiple mappings. Configuration is available via API, console, CLI, SDKs, CloudFormation, or SAM, with costs tied to the Event Poller Unit (EPU) billing model. For artifact delivery, Amazon ECR now supports image layers up to 200 GB for images pushed via Docker, addressing prior limits that forced splitting or external storage for large assets such as embedded language models, genomics datasets, or sizable binaries. Workflows using the UploadLayerPartAPI via SDK/CLI remain limited to 50 GB per layer, and the increase excludes the Middle East (Bahrain) and Middle East (UAE) Regions.

EC2 I7i instances expanded to Asia Pacific (Thailand) and Israel (Tel Aviv). Powered by 5th Gen Intel Xeon processors (3.2 GHz all‑core turbo) and 3rd‑generation AWS Nitro SSDs, the I7i family targets storage‑optimized, I/O‑intensive, and latency‑sensitive workloads. AWS cites up to 23% better compute performance and more than 10% better price performance versus I4i, with up to 45 TB of NVMe, up to 50% better real‑time storage performance, up to 50% lower I/O latency, and up to 60% lower latency variability. Torn write prevention supports block sizes up to 16 KB, and sizes include nine virtual up to 48xlarge and two bare metal options, with up to 100 Gbps network and 60 Gbps EBS bandwidth.

Active Threats and Patch Guidance

Unit 42 detailed three techniques—Pass‑ta‑key, Silver Pass‑ta‑key, and Golden Pass‑ta‑key—that enable malware on a Windows device to abuse Google‑synced passkeys in Chrome. The attacks do not break passkey cryptography but exploit handling of device identity keys, onboarding, recovery, and synced credential state. Pass‑ta‑key can impersonate a trusted device using a TPM‑backed identity key to obtain a signed assertion from the cloud authenticator, which may work if the relying service does not strictly enforce the User Verified flag. Silver Pass‑ta‑key can register an attacker’s user‑verification key during forced re‑registration, and Golden Pass‑ta‑key abuses exposure of the security domain secret in process memory to decrypt synced passkey records. Recommendations include enforcing user verification at relying parties, hardening recovery and re‑registration flows, and preventing exposure of master secrets in memory.

BleepingComputer reported that N‑able disclosed active exploitation of an authentication bypass in N‑central (CVE‑2026‑18577). Hotfix 2026.3.1.7 is available, with hosted instances patched by the vendor and on‑premises deployments requiring manual updates. The issue relates to an incomplete fix for a prior flaw (CVE‑2026‑18576) and could enable administrative account takeover. Indicators of compromise include four IP addresses, a registered service named “Cloudflared,” and the presence of “svchost.exe” in users’ Documents folders. N‑able advised immediate upgrades, close monitoring, and contacting support if IOCs are detected.

The Hacker News covered three high‑severity vulnerabilities (FaceHugger) in Hugging Face’s Diffusers library that enable arbitrary code execution when loading models from crafted Hub repositories. The TOCTOU issues permit code injection via the custom_pipeline mechanism even when trust_remote_code is False (CVE‑2026‑44827, CVE‑2026‑44513; CVSS 8.8) and exploit a race between hf_hub_download and snapshot_download (CVE‑2026‑45804; CVSS 7.5). Maintainers addressed the flaws in Diffusers 0.38.0 (early May 2026) and recommend immediate patching, with interim mitigations to treat model repositories as untrusted code sources, avoid unaudited custom pipelines, and inspect local snapshots for unexpected .py files before invoking from_pretrained.

The Hacker News also reported a Thermo Fisher Scientific update for CVE‑2026‑17583 affecting Applied Biosystems human identification products. The vendor rated the issue High (CVSS v4.0 8.2) and added digital signatures so future .fsa and .hid files can be validated; three end‑of‑life data collection products will not be updated. Thermo Fisher recommends immediate installation of fixes and, where updates are not possible, compensating controls such as strict chain‑of‑custody, encrypted and access‑restricted storage, least‑privilege access, and limited network connectivity. The company stated it is unaware of exploitation in the wild, and the CVE was credited alongside CISA.