< ciso
brief />
Cloud Governance Upgrades, Agentic AI Oversight, and Critical Patches

Cloud Governance Upgrades, Agentic AI Oversight, and Critical Patches

Coverage: 13 Nov 2026 – 15 Nov 2026 (UTC)

< view all daily briefs >

Cloud and AI platforms introduced new governance and security controls while researchers detailed AI-enabled offensive tooling and several urgent fixes. Highlights include organization-wide posture management, cross-account secret replication, multicloud systems management, and tighter Kubernetes job security. On the risk side, an actively exploited UnrealIRCd flaw, large patch sets for PostgreSQL, and new attacks against AMD SEV-SNP underline the need for prompt remediation. Organizations operating AI agents also gained observability options as defenders adapt to autonomous orchestrators.

Centralized Cloud Security and Identity Controls

AWS Security Hub Cloud Security Posture Management (CSPM) now supports declarative policies managed through AWS Organizations, enabling administrators to configure posture checks across all accounts and Regions—or specific account–Region combinations—and to automatically apply baselines to new accounts. Declarative policies are recommended over prior central configuration to reduce drift and enable Region-specific settings. A built-in migration in the CSPM console converts existing central policies to the new format, and the feature is available across all commercial, AWS GovCloud (US), and China Regions.

AWS Secrets Manager introduced cross-account replication, allowing synchronized copies of a primary secret to be maintained in one or more secondary accounts for access, backup, and disaster recovery. Replicas are independent resources encrypted with KMS keys owned by the replica account, eliminating the need for shared keys or cross-account policies. Secrets values propagate automatically on change or rotation, and replicas can be promoted to standalone secrets during recovery. The capability extends existing multi-Region replication and is available wherever Secrets Manager is offered.

Systems Manager expanded its Cloud Connector to include Google Cloud, automating SSM Agent deployment and presenting Google Cloud VMs alongside EC2, on‑premises, and Azure resources in one view. Teams can apply familiar capabilities—Session Manager, Automation runbooks, Run Command, State Manager, Patch Manager, and Inventory—across a multicloud estate to standardize operations, reduce configuration divergence, and streamline incident response across providers.

AWS Batch on Amazon EKS added support for container-level Linux capabilities, seccomp profiles, and pod-level settings such as fsGroup. These options, configurable via RegisterJobDefinition and the container or pod securityContext, help organizations meet the Kubernetes Pod Security Standards restricted profile by minimizing privileges and constraining system calls. The controls are accessible through the console, CLI, SDKs, and CloudFormation in all Regions where AWS Batch is available.

AI Operations: Observability and Orchestrators

BigQuery and Looker now provide expanded AI agent observability and evaluation across Google’s Data Cloud. The update centralizes telemetry from first‑party Conversational Analytics agents and custom agents (e.g., ADK or LangGraph) into customer‑owned BigQuery tables, with operational logs governed by Cloud Audit Logs, VPC Service Controls, and customer‑managed encryption keys. Teams can enable Cloud Monitoring and Cloud Trace, opt in to per‑agent conversation logging with privacy guardrails, instrument tool‑invocation traces via an open‑source SDK plugin, and use BigQuery AI for root-cause and semantic drift analysis. Cost controls include token and latency tracking, prebuilt dashboards in BigQuery Studio and Looker System Activity, and quotas such as per‑query byte‑scan caps and daily token limits. Integrated evaluation via Prism and BigQuery APIs supports benchmarking and side‑by‑side comparisons with production logs.

Cisco Talos reported broader misuse of AI-enabled penetration testing frameworks that chain reconnaissance, exploitation, and post‑exploitation with minimal operator input. ARTEX, released in July 2026, uses planner/worker agents, integrates multiple LLMs, and connects to tool servers via MCP; Talos observed adoption by financially motivated groups and actors with a China nexus, spanning scanning, credential theft, and targeted intrusions against multiple sectors. Techniques include password spraying, exploitation of web and authorization flaws, cloud abuse, privilege escalation, and deployment of web shells and third‑party C2. Talos enumerated over 2,500 ARTEX‑associated C2 IPs and provided Snort and ClamAV coverage, advising defenders to reduce exposure, harden credentials and cloud identities, and detect agent‑like automation.

Cloudflare report documented a rapid rise in LLM‑based penetration testing orchestrators, with ARTEX‑derived projects forming a majority and a “hydra” effect after the original repo’s takedown. Incident response cases included lateral abuse of static CI/CD tokens and unrestricted API keys, leading to tenant access and data exfiltration. Cloudflare recommends a Modern Security Model emphasizing perimeter traffic control, WAF and rate‑limiting, bot defense, zero trust access and micro‑segmentation, strict API and ingress validation, and intelligence‑driven detection. The guidance stresses least privilege, curbing long‑lived credentials, and containment‑first architectures to slow autonomous attack chains.

Active Exploits and Patch Imperatives

UnrealIRCd flaw (CVE‑2026‑114713, CVSS 8.6) was added to CISA’s Known Exploited Vulnerabilities catalog amid confirmed exploitation. The issue in the labeled‑response module fails to escape the IRCv3 ‘label’ message tag, allowing users on multi‑server networks to inject server‑to‑server commands with server authority. UnrealIRCd 6.2.7.1, released November 7, 2026, addresses this and other issues; project lead Bram Matthys urged immediate hotfixing or, at minimum, a spamfilter workaround. Under BOD 26‑04, U.S. Federal Civilian Executive Branch agencies face a November 16, 2026 remediation deadline.

Microsoft 365 outage recovery is in progress with stabilization signs, but security teams are warned of elevated phishing, fake support calls, and other social engineering during the disruption. Authorities urged caution around unsolicited contact, and experts advised treating the period as an incident—reviewing sign‑ins, mailbox rules, account changes, and endpoint activity originating during monitoring gaps, and gathering evidence directly from devices rather than solely the affected platform.

PostgreSQL updates closed 52 flaws, including CVE‑2026‑6460 (CVSS 8.8), a type‑confusion issue that can let authenticated users who can create objects run code with the database process’s OS privileges. Fixes add version checks to detect mismatched composite types. The project recommends upgrading to 18.7, 17.12, 16.16, 15.20, or 14.25, and rebuilding all C‑language extensions to avoid residual exposure; some extensions may require source changes. Client tools (pg_dump, pg_restore, reindexdb, vacuumdb) also received fixes. PostgreSQL 14 is now end‑of‑life, with 14.25 the final release.

AMD SEV-SNP confidential VMs face two new attacks: Mior and Skyfall. Mior corrupts the Reverse Map Table during startup by redirecting Platform Security Processor writes to a controlled PCIe device, enabling a malicious hypervisor to rewrite RMP entries, enable debugging, and read/modify VM memory and falsify attestation—even with ciphertext hiding. Skyfall exploits Transient Scheduler Attacks and cached VM state after failed startups to recover registers and a disk‑encryption key and leak specific cache‑line offsets. AMD assigned CVEs and released SEV firmware and microcode updates; platform firmware and vendor BIOS updates are required, with some embedded fixes extending into 2027.

Infrastructure for HPC, AI, and Databases

Google Cloud HPC showcased agentic technologies at SC26, packaging tested HPC and scientific expertise into reusable agent skills integrated with Cluster Toolkit and available via Google Skills and DeepMind Science Skills repositories. Skills handle quota and capacity checks, workload‑to‑blueprint matching, IAM and budget guardrails, and Day‑2 diagnostics. DeepMind’s AlphaEvolve, powered by Gemini and running with Cloud Batch, executes large‑scale evolutionary loops to mutate, benchmark, and optimize code and 3D parameters across hundreds of concurrent evaluations on real fabrics using Cloud RDMA and Managed Lustre, with cost controls including Spot VMs and DWS. Google also detailed purpose‑built infrastructure managed by Cluster Director—memory‑optimized H4D and Arm‑based C4A instances, multi‑node GPU A4/A4X and A5X bare metal on NVIDIA Vera Rubin platforms, G4 RTX PRO instances for CAE/CFD, 8th‑gen TPUs (TPU 8t and 8i) with Batch support, and sub‑millisecond storage via Managed Lustre with GPUDirect over RDMA and an 80 PB Dynamic Tier.

Aurora PostgreSQL with express configuration now supports AWS CloudFormation via the withExpressConfiguration createOnly property on AWS::RDS::DBCluster. Templates can provision the cluster and its primary instance with a single property, shortening definitions and reducing operational overhead. Clusters created this way can be queried within seconds, supporting rapid serverless workflows; the capability is available in all Regions where express configuration is offered.

Grok 4.7 is now available in the Kiro IDE and CLI for AWS GovCloud (US), described as SpaceXSI’s most capable model for coding and long‑running knowledge work, with improved planning, error recovery, and an experimental 500K token context window. Billing applies a 2.7x credit multiplier up to 200K tokens and 5.4x beyond. Kiro administrators can control model availability to align with organizational compliance and policy.