Cloud Security Updates, AI Stack Advances, and Critical Git Fix

Cloud providers announced security-focused features alongside broader AI and data platform updates. Amazon Web Services detailed new controls and observability across networking, storage, databases, and AI services, while Google introduced governed metrics modeling, a healthcare agent, secure code-fixing automation, and an open AI-native Kubernetes stack. The Git project released fixes for a critical vulnerability, and separate incidents impacted industrial refrigeration in France and campus operations at Vanderbilt University.

AI Platforms, GPUs, and Agent Capabilities

AWS G7 GPU instances expanded to new Regions including London, Mumbai, Sydney, and São Paulo. Powered by NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs and custom Intel Xeon 6 processors, G7 offers up to eight GPUs (32 GB per GPU), 192 vCPUs, 768 GiB memory, 7.6 TB local NVMe, and up to 700 Gbps EFA networking. AWS cites up to 4.6x AI inference and 2.1x graphics performance improvements over G6 for workloads spanning AI inference, image/video analysis, speech recognition, recommender systems, real-time rendering, game streaming, spatial computing, and video transcoding. Instances are available On-Demand, as Spot, or via Savings Plans through the console, CLI, or SDKs.

Bedrock Daybreak adds access to OpenAI’s Daybreak Blue: GPT-5.6 Sol for defensive security workflows and Daybreak Red: GPT-5.6 Cyber for authorized offensive research. Access requires enrollment with OpenAI followed by AWS approval; AWS offers US and Global cross-Region inference profiles. OpenAI and AWS state that inference data is not used for model training, and customers are not required to opt in to data sharing.

Gemini healthcare launched in preview in the U.S. as an agent to orchestrate multi-step clinical, administrative, and payer workflows. It features reusable skills, secure connectors to EHRs and operational systems, an open partner ecosystem, and a platform with enterprise access controls, audit logging, and verifiable citations. The agent enforces existing identity and role-based permissions, supports HIPAA compliance, and allows model choice under unified governance, with clinical data kept within the organization’s governed environment.

AI-native stack guidance from Google describes an open-source Kubernetes stack in the CNCF Sandbox for deploying RL-trained and open-weight models. Components include OpenRL for post-training time-slicing, llm-d 1.0 for prefix-cache-aware routing and multi-tier KV cache offload, KubeRay and Kueue for orchestration and scheduling, Agent Substrate for agents, and gVisor for sandboxing. Reported benchmarks show multi-step reasoning gains with dynamic TPU/GPU sharing, and sample manifests are available for trying the stack on Kubernetes and GKE.

AgentCore Search is generally available, providing a Web Search API that returns ranked results from an Amazon-operated web index with text extracts, titles, URLs, and timestamps. Developers can filter by domain and time, integrate via AWS SDKs and CLI with IAM authorization, route traffic via PrivateLink, and log searches as CloudTrail data events. Pricing is $7 per 1,000 queries, initially in selected Regions including N. Virginia, Ohio, Oregon, Sydney, Tokyo, and Ireland.

Resilience and Governance in Data and Analytics

OpenSearch DR introduced managed cross-Region disaster recovery. Customers pair primary and secondary domains via API or console; the secondary remains continuously synchronized. The service targets an RPO under one minute and RTO under 20 minutes, supports switchover with zero data loss, and enables failover with data loss limited to replication lag. Health and lag metrics surface in CloudWatch. Requirements include OpenSearch 3.7 or later, with certain storage and instance types not supported; there is no separate DR charge beyond secondary domain and data transfer.

OpenSearch upgrades now support clone-based version migrations that clone a target-version domain while the primary continues serving workloads. Operators can validate queries and performance side-by-side via CloudWatch, cancel prior to cutover, and roll back during a retention window after cutover. The feature supports domains on OpenSearch 1.1+ or Elasticsearch 7.10 in most Commercial Regions, with noted exceptions.

BigQuery measures is generally available, unifying governed BI metrics with knowledge graph relationships. The GA adds composite and derived measures, compile-time validation to catch circular references, and a start_node argument in GRAPH_EXPAND to avoid chasm traps across multiple fact tables. Rich semantic metadata flows into Knowledge Catalog, and integrations span Data Studio, LookML reuse, and authoring support through Data Agent Kit.

S3 Access Grants now supports cross-account grant discovery with a single API call by setting InstanceScope=ALL in ListCallerAccessGrants. The same behavior applies to GetDataAccess, which locates the owning account’s instance and returns credentials for the requested path. The capability is available in all supported Regions at no extra cost and reduces operational overhead in multi-account setups.

S3 Tables logging enables S3 server access logs for table buckets, delivering request-level records to a designated bucket in the same account and Region. Logs capture requester identity, operation, table, namespace, response status, and timing in the familiar space-delimited S3 log format, supporting audits, analytics, and cost attribution via tools such as Amazon Athena or DuckDB. There is no additional delivery fee; customers pay for log storage.

CloudWatch PromQL alarms can now trigger scaling actions for EC2 Auto Scaling groups and Application Auto Scaling, using simple or step scaling policies. Queries must return a single time series against a threshold; target-tracking policies are not supported at this time. The capability is available in all Regions that already support PromQL alarms, with standard CloudWatch alarm pricing.

Databases and Developer Workflow Updates

RDS Postgres ES offers Extended Support for minor versions 13.23-rds.20260819, 12.22-rds.20260820, and 11.22-rds.20260822. The program provides up to three additional years of fixes for critical CVEs and significant bugs beyond standard end-of-support, giving organizations more time to plan major upgrades. Customers can use Blue/Green Deployments, in-place upgrades, or restore from snapshots via the console or CLI.

RDS MySQL 26.10 is available in the RDS Database Preview Environment. The pre-release includes bug fixes, security patches, and performance improvements for READ COMMITTED workloads by reusing read-view objects to remove a global mutex. Observability gains include error-log warnings when history-list length exceeds innodb_max_purge_lag or when purge lag delays DML. Preview instances are retained up to 60 days, with snapshots usable only within the Preview Environment; pricing aligns with US East (Ohio).

CodeMender integrates an autonomous code-security agent into pull requests on Google Cloud Run. It posts a single comment with findings and supports /verify and /fix to simulate exploits and author patches; /approve can commit successful fixes and set a security gate status. Verification artifacts are archived to Cloud Storage, and security controls include short-lived OIDC tokens, immutable repository identifiers, minimal Cloud Run job privileges, and no network egress by default.

Mythos 5.2 and Claude Fable 5.2 were introduced with faster output, a 20% API price cut, improved coding performance, and better vulnerability rediscovery. Fable 5.2 scored 49.1% on FrontierCode 2.0 (Xhigh) versus 32.4% for Fable 5.1; Mythos 5.2 identified 89.6% of issues in an internal CVE rediscovery test. Access models differ—Fable is broadly available, while Mythos access is limited to vetted cybersecurity and life-sciences organizations. Eligible customers can opt for zero data retention; others face 30-day storage for safety monitoring.

Security Controls, Patches, and Incidents

Network Firewall NGFW policy separates access-control rules from intrusion prevention system rules, increasing effective capacity and simplifying management. Teams can enforce deny-by-default firewall rules and deploy IPS protections using managed or custom rule groups. Embedded rule groups draw capacity on demand, all rules operate statefully, and existing IPS rules remain backward compatible. The capability is available at no additional charge across all commercial AWS Regions and AWS GovCloud (US).

DocumentDB 8.0.3 adds hybrid post-quantum TLS for TLS 1.3 using ML-KEM when supported by clients, with CloudWatch metrics distinguishing handshake types. Developer compatibility expands with wildcard indexes, $unionWith in aggregation pipelines, $lookup and $graphLookup in views, and change stream events for dropIndexes operations. Storage and performance visibility improves via collStats and stats() reporting. The release is available in all Regions where the service is offered.

Git flaw CVE-2026-111760 was patched to prevent arbitrary code execution during recursive clones of malicious repositories with submodules. The issue involved an overly long path passing safety checks but truncating on disk to attacker-controlled names like .git, enabling symlink-based hook execution. Fixes landed in Git 2.56.1 with backports for 2.50–2.55, and additional vulnerabilities were addressed across protocol handling, GPG-signed tag merging, oversized binary patches, large-merge buffers, and git-fast-import features. Users are advised to update Git and avoid recursively cloning or merging untrusted sources.

France OT attack attribution cites an IRGC-linked unit for destructive actions against retail refrigeration systems. On October 11, attackers reportedly altered defrost settings across around 700 sites, leading to at least 15 fires and significant losses; affected controllers were described as older Danfoss models. Authorities tied activity to Shahid Kaveh based on scanning reuse, Farsi traces, and rapid propaganda claims, and named three individuals allegedly linked to IRGC cyber leadership. A second wave in early November affected businesses across Europe, including blood services, without formal attribution. Guidance emphasizes securing OT by removing internet exposure, updating software, changing defaults, and segmenting and monitoring networks.

Vanderbilt incident prompted system shutdowns on November 9, with users told to power down university devices and disconnect from campus Wi‑Fi. An evening update said classes and student services would run normally from November 10, Wi‑Fi was available, but most systems remained restricted during investigation with government and external partners. The BYOD ransomware group claimed responsibility, alleging access to extensive data including research, over one million social security numbers, and control of the domain controller; Vanderbilt has not confirmed technical details and the response is ongoing.