Cloud Platforms Expand Serverless Speed as Critical Flaws Emerge

Cloud platforms delivered a wave of performance and operations updates spanning serverless cold-starts, Kubernetes traffic management, observability, and recovery testing. At the same time, multiple critical vulnerabilities drew urgent attention, including actively probed Oracle E‑Business Suite and newly KEV-listed Adobe Commerce and Brotli issues. Vendor research also highlighted AI’s growing role on both sides of the security ledger—from tools that streamline agent development and approvals to demonstrations of automated exploit chaining.

Faster Startup and Smarter Traffic for Cloud-Native Workloads

AWS Lambda added Seekable OCI (SOCI) lazy loading for Lambda Managed Instances, enabling Event Functions packaged as OCI images to fetch only the files needed during environment startup. AWS cites up to 70% reductions in container startup times for multi‑gigabyte images. Adopting SOCI requires generating a SOCI index, pushing it to Amazon ECR, and referencing the indexed tag; Lambda detects the index automatically. The feature is available in all commercial Regions where Lambda Managed Instances is supported.

EKS Auto Mode now supports the Kubernetes Gateway API, offering native resources such as Gateway and HTTPRoute to define how external and internal traffic reaches workloads. Capabilities include HTTP/gRPC routing, cross‑namespace routing, weighted splits for canary or blue/green deployments, header‑based routing, and automatic TLS certificate discovery and provisioning via AWS Certificate Manager. Existing Ingress configurations remain compatible, and EKS Auto Mode provisions AWS load balancing infrastructure automatically.

GKE Networking guidance detailed how Dataplane V2 scales highly concurrent agentic AI workloads by separating high‑cardinality identity data from functional metadata and using Cilium label exclusion filters. By moving non‑selector fields from Pod Labels to Annotations and excluding dynamic labels, many identical agent pods can share a single numeric network identity, avoiding limits and preserving eBPF‑based enforcement. Google cited validations including hundreds of thousands of secure environments and sustained activations with sub‑500 ms suspend/resume latency across 200,000 agents.

Observability, Restore Readiness, and Data Lifecycle at Scale

OpenSearch Serverless introduced time series collections with a new warm storage tier alongside hot, enabling lower‑cost retention for infrequently accessed data while maintaining performance for recent logs, metrics, traces, and events. Each collection supports up to 100 TB of hot storage and can absorb ingestion spikes. Integrated alerting and anomaly detection help surface issues directly from observability data, with availability in all Regions where OpenSearch Serverless is offered.

Lambda Telemetry for durable functions adds a Durable Execution Telemetry Exporter that captures execution histories—status, operation timing, step I/O, and errors—and forwards them to destinations including Amazon CloudWatch Logs, Amazon S3 (for Amazon Athena), Amazon DynamoDB, Amazon Aurora, and Amazon OpenSearch Service. Operators can redact sensitive fields and configure sampling to control data volume and retention. A companion Model Context Protocol server allows AI agents to query telemetry with natural language.

AWS Backup added restore testing for Amazon Redshift, allowing automated, scheduled restore tests that select eligible recovery points, perform restores, and record completion times against defined RTOs. Resources can be targeted by ARN or tags, tests may restore the latest or a random recovery point, and an optional validation step verifies the restored resource before temporary assets are removed. The capability is available wherever restore testing and Redshift are supported.

Cloud SQL announced managed blue‑green deployments in preview for Cloud SQL for MySQL, creating a synchronized green environment to validate upgrades using production data before a near‑zero‑downtime switchover. During cutover, the blue instance is briefly set to read‑only while final sync completes; IP, instance name, and backups transfer to the green instance, avoiding application changes. Rollback is possible by deleting the green instance before switchover.

Agentic AI: Tooling for Builders, Guardrails for Actions

AgentCore CLI version 1.0 is generally available for Amazon Bedrock AgentCore, delivering a refactored, faster CLI with a consistent command structure and expanded support for resources like harnesses, runtimes, memory stores, identity, gateways, and payments. New capabilities include invoking resources by ID/ARN, an interactive shell, harness export, and guided, searchable workflows. A migration command helps upgrade existing projects, and the tool is distributed via npm with source on GitHub.

Amazon Quick added smart approvals for agentic actions, automatically approving routine, low‑risk actions while sending higher‑impact or out‑of‑scope actions back to users for review when the user’s setting is set to “ask.” The feature respects existing policies, cannot grant new permissions, and records all reviews and setting changes in AWS CloudTrail. Administrators control access via permissions, and users can disable smart approvals per conversation or scheduled run.

Active Threats, Urgent Patches, and Emerging Techniques

Oracle EBS customers face active probing of CVE‑2026‑101767, a critical unauthenticated SQL injection in the Oracle Application Object Library (TCF servlet) that can be chained to remote code execution as the EBS application account. Researchers recommend immediate patching and, where not possible, restricting access paths and monitoring for indicators such as unexpected files in $APPLPTMP and shells spawned by the oracle account.

Adobe patches address eight critical issues in Adobe Campaign Classic (build 9403) and a separate critical code injection flaw (CVE‑2026‑48360) in Adobe Connect Desktop and Mobile, enabling arbitrary code execution. Adobe is not aware of active exploitation but urges customers to update and review exposure and access controls.

PeopleSoft flaw exploited by ShinyHunters (UNC6240) comprises a chain of three issues tracked collectively as CVE‑2026‑35273, beginning with SSRF in the PeopleSoft Integration Gateway and leading to internal admin access and remote code execution via web shell or XMLDecoder paths. Oracle patched the chain in June; guidance includes applying updates, disabling or removing PSEMHUB if necessary, and treating web‑shell hosts as fully compromised.

Adobe Commerce PolyShell (CVE‑2026‑48356) and Google Brotli decoder (CVE‑2026‑111497) were added to CISA’s KEV catalog amid active exploitation attempts. Attacks on Adobe Commerce/Magento have breached hundreds of stores via unrestricted file uploads, while public tooling targets vulnerable Node.js Brotli decoding to achieve shell access. Vendors across ecosystems have issued patches; federal agencies face a November 10, 2026 mitigation deadline.

EDK II PXE contains a 16‑year‑old buffer overflow (CVE‑2026‑97631) in its network‑boot path, allowing crashes or potential code execution inside firmware before Secure Boot initializes. TianoCore merged a fix, and Microsoft’s Project Mu integrated it, but many edk2 packages lack released updates; operators should disable PXE where not needed and monitor boot‑time network activity pending firmware releases.

Tailscale SSH 1.104.4 fixes several issues that could let an attacker controlling a permitted device reach services and files restricted to root via socket forwarding. The fix changes forwarding to run as the logged‑in user and defers access control to OS permissions; administrators can update or disable SSH forwarding and review access rules while patching.

PHP2Shell attack disclosures describe a coordinated intrusion that rooted 25 internet‑facing firewall systems across multiple countries using custom exploits for PHP2Shell RCE bugs (CVE‑2026‑109618/109619) combined with a Linux kernel privilege escalation (CVE‑2026‑46300). The actor established persistence and attempted Active Directory attacks; many compromised devices reportedly retain active web shells.

GMO Internet suspended its Onamae.com virtual desktop service after detecting malicious software in customer server environments. Up to 26,739 servers may be impacted. Following an extortion demand and ransomware threat, the company advised backups, virus scans, and disconnecting machines; forensic analysis continues.

Anthropic research described an internal model that located and chained multiple vulnerabilities into working proofs of concept across Chrome, Safari, Android, Windows, and Linux in a lab setting. Issues were disclosed to vendors and patched; Anthropic emphasized the significance of combining separate flaws with limited human involvement.

Skynet agent analysis by Symantec documented an offensive AI agent that hijacks exposed model endpoints with multilayered jailbreak prompts to conduct reconnaissance, data theft, remote code execution, and other actions while shifting compute costs to victims. Recommendations include authenticating all model endpoints, monitoring inference spend, and enforcing least privilege with human approval for sensitive actions.

Android Stores gained a new Registered App Stores program outside the U.S., allowing qualifying third‑party app stores to perform one‑tap installs and background updates on Android 17 QPR2+ devices in supported countries. Stores must submit catalogs for Google malware scanning, pay an annual fee, and meet defined thresholds; noncompliance can lead to suspension. Users must consent to installations, and visible notifications are required for large/background downloads.