Cloud Security Automation, Agentic Runtimes, and PHP2Shell Response

Cloud providers emphasized automation and governance across security and AI stacks, while defenders moved quickly to blunt active exploitation of new PHP engine flaws. Network and compute expansions rounded out the day with new options for private connectivity, regional availability, and reliable agent execution. The developments collectively target faster remediation, stronger isolation, and consistent governance at scale.

Automation in Vulnerability Management and Resilience

AWS Continuum for code vulnerabilities is now generally available as a managed service that automates the end-to-end handling of code risks. It builds a context graph spanning resource topology, identity relationships, and network reachability to identify findings that matter to applications and business risk. Continuum ingests existing vulnerability data, performs code scans, and validates exploitability in an isolated sandbox. It prioritizes by impact and generates remediation patches as pull requests, complete with proof of exploitability and blast-radius assessment. Multiple AI models power the workflow, which starts in Learn mode with human approval and can progress to Enforce mode for autonomous remediation within guardrails. The service is designed to complement Amazon GuardDuty and AWS Security Hub and follows standard AWS regional availability.

CloudFormation added a force rollback option to return stacks to a known good state after failed updates. Instead of retrying resource operations individually, administrators can complete rollback in a single command, with any resources requiring manual intervention recorded for review. The capability integrates across the AWS CLI, SDKs, AWS CDK, and AI-driven workflows. AWS recommends running continue update rollback with the force flag, then using a drift-aware change set to inspect differences before redeploying. The feature is available in all Regions at no additional cost and aims to reduce remediation time and complexity in CI/CD pipelines when partial failures occur.

Active Exploitation: PHP2Shell and Orchestrator Risks

Check Point Research reports that exploitation attempts against the PHP2Shell vulnerabilities (CVE-2026-109618 and CVE-2026-109619) are impacting about 10% of organizations globally, with Europe hardest hit at 14% and peak activity around midnight. Telecommunications and education are the most affected sectors. Because the flaws reside in PHP itself, updating CMSs or plugins is insufficient; servers must be upgraded to fixed PHP releases (8.2.35, 8.3.36, 8.4.27, 8.5.12) or have mitigations backported. The guidance stresses full inventory of PHP instances (including containers and bundled runtimes), prioritizing internet-facing systems, restarting services, and verifying versions, alongside monitoring for anomalies in files, web requests, and outbound connections. Check Point notes that leaked details and tooling accelerated exploitation.

As a defense-in-depth measure during patch windows, AWS WAF guidance outlines edge-layer rules to reduce exposure to confirmed attack paths. Recommended controls center on multipart/form-data traffic, the presence of a PHP session identifier, and the PHP_SESSION_UPLOAD_PROGRESS field, with rules first deployed in Count mode and then moved to Block after validation. Additional steps include body-size caps and rate limits. The post reiterates that upgrading to PHP 8.5.12, 8.4.27, 8.3.36, or 8.2.35 is the only complete remediation and lists host-level mitigations (for example, disabling session.upload_progress and setting appropriate memory limits) to apply in parallel.

Consumer environments are also at risk. Kaspersky warns that PHP2Shell can affect home NAS, smart-home controllers, surveillance systems, gaming servers, and personal cloud installations. Several vendors (e.g., Jeedom, QNAP, Synology, ASUSTOR, Unraid) issued guidance or updates; PHP 7.3 will not be fixed and PHP 7.4+ require vendor updates. Recommended actions include installing all available patches, disconnecting vulnerable devices from the internet or restricting them to VPN/local access, backing up data, removing outdated PHP packages, and updating containerized/bundled PHP instances individually. For devices unlikely to receive fixes, the advice is to keep them offline or internal-only until remediation is possible.

Separately, Arista addressed an actively exploited critical SQL injection in VeloCloud Orchestrator (CVE-2026-114625) that allows an authenticated user with basic privileges to escalate to operator or superuser without further interaction. Hosted environments have been patched and on-premises fixes are available for specific versions (5.2.3.18+, 6.1.3.10+, 6.4.2.12+). Interim mitigations include restricting access to trusted administrative networks, preserving and reviewing logs, monitoring for unusual request patterns, and blocking IPs observed in attacks. BleepingComputer notes that this follows several other exploited issues earlier in the year, underscoring risks when authenticated attackers can leverage credentialed access in orchestration layers.

Agentic AI Runtimes, Context, and Reliability

Google introduced Agent Compute Engine and expanded governance features for Gemini agents to support large-scale autonomous workloads. The platform provides purpose-built, high-density microVMs with rapid boot, copy-on-write memory sharing, fast checkpoint/restore, hardware isolation via Titanium IPU, per-agent identities, and persistent Hyperdisks for workspace continuity. For Kubernetes users, Agent Substrate on GKE offers an open-source, zero-trust runtime with sub-second suspend/resume and portability. Governance additions—Agent Gateway, Agent Registry, SPIFFE-based Agent Identity, and OpenTelemetry-aligned observability—extend policy enforcement and auditability across Google Cloud, hybrid/multi-cloud via Envoy-based points, and local runtimes.

To improve AI agent context quality, Google announced preview releases of self-learning context enrichment and a redesigned Knowledge Studio in Knowledge Catalog. Gemini-powered enrichment analyzes agent/user interactions to detect missing or outdated context, ingests verified definitions and metrics from sources such as playbooks, GitHub, Confluence, Google Workspace, and structured metadata, and proposes updates with citations and confidence scores. Changes can auto-publish above a threshold or route as Change Requests for steward approval. The new Knowledge Studio surfaces a visual context graph and workflows, while governed context sets bind domains or agents to the right segments. Google cites internal benchmarks showing improvements in accuracy, latency, and token usage, and states it does not use customer data to train its models.

On AWS, new SageMaker Sessions support serverless model customization with managed techniques such as supervised fine-tuning (SFT) and DPO, while exposing low-level operations (e.g., forward_backward(), optim_step(), sample()) for bespoke training and reinforcement learning. SageMaker manages GPU allocation, LoRA/model loading, distributed execution, checkpoints, and fault recovery. Sessions are billed token-based, excluding local computations such as reward scoring, and are available initially in US East (N. Virginia), US West (Oregon), and Asia Pacific (Mumbai), with a published cookbook to accelerate adoption.

AgentCore Gateway now supports the InvokeModel and InvokeModelWithResponseStream APIs on its inference endpoint and adds ordered model fallback with circuit breaking. Administrators can trigger fallback on throttling, server errors, or timeouts, and the gateway returns a single response even if multiple routes are tried internally. This reduces client-side retry logic, simplifies multi-provider deployments, and improves availability across supported Regions (excluding China).

AWS also expanded serverless customization to new models—Muse Glimmer, Qwen3.8 27B, and Ministral 3 (3B, 8B, 14B)—enabling SFT, DPO, and reinforcement fine-tuning with LoRA or full-parameter approaches. The feature is intended to produce smaller, cost- and latency-optimized models aligned to task-specific workloads, with regional availability presently in N. Virginia, Oregon, Tokyo, and Ireland. Teams can launch jobs from SageMaker Studio, the Python SDK, or agent skills. SageMaker Customization complements an existing catalog from families such as Nova, Gemma 4, Nemotron 3.5, Qwen, Llama, gpt-oss, and DeepSeek.

Networking and Regional Availability

AWS Interconnect – last mile entered preview with Megaport in the San Jose and N. Virginia metros, enabling private, high-speed connections from branches and data centers to AWS using cloud-native provisioning. The service pre-provisions capacity and automates BGP peering, VLAN configuration, and ASN assignment, with plans to scale bandwidth dynamically from 1 Gbps to 100 Gbps at GA and a 99.99% SLA. Megaport contributes a global SDN footprint and an open API for partner adoption.

Direct Connect expanded 100 Gbps capacity at the Netmagic DC2 site near Bangalore, enabling private links to all public AWS Regions (except China), to AWS GovCloud, and to Local Zones. The upgrade targets higher throughput and consistent performance for hybrid architectures and data-intensive applications, adding to a portfolio of more than 150 Direct Connect locations.

Lambda MicroVMs added four Regions—Hong Kong, Malaysia, Thailand, and Spain—bringing total coverage to 14. MicroVMs deliver VM-level isolation with near-instant launch and resume times and state preservation, supporting use cases such as AI coding assistants, interactive development environments, and vulnerability scanners. Developers build images from Dockerfiles and launch instances with dedicated HTTPS URLs supporting HTTP/2, gRPC, and WebSockets.

Amazon Quick is now available in the AWS Singapore Region, offering an agentic AI teammate that answers questions, performs research, organizes work into Spaces, and surfaces insights via QuickSight dashboards. Local deployment supports data sovereignty and compliance needs for organizations operating in Singapore.

Google Cloud expanded deployment choice for Anthropic models, adding an Anthropic-managed option alongside the existing Google-managed path. The Anthropic-managed route provides day-zero access to platform features like Managed Agents (beta), while keeping Anthropic as data processor; the Google-managed approach offers additional governance controls including VPC Service Controls, CMEK, and regional data residency across U.S., EU, and APAC. Both share pricing, unify billing and IAM, and provide access to the latest Claude models. Claude on GCP is positioned to help organizations meet compliance and sovereignty requirements without changing procurement or identity flows.