Weekly ML Roundup: Model Backdoors, Azure Compute and Local AI
New model-security research shows that poisoned post-training data and edits to fewer than ten neurons can implant hidden behavior that ordinary task evaluations miss, making checkpoint audits and backdoor-specific testing important before deployment. Azure opened HDv2 preview registration for data preparation and dense agent sandboxes, while GPU inference tests reported performance in line with reference benchmarks, with preliminary sandbox estimates still requiring workload-specific validation. Windows ML introduced experimental llama.cpp support and task-specific text generation and speech recognition APIs, giving developers another route to local inference without replacing existing ONNX Runtime integrations. OneLake storage support for Unity Catalog reached general availability in Fabric while remaining in preview on Azure Databricks, and Fabric Warehouse previewed Git-based schema deployment alongside an upcoming SQL pool API change that requires client updates. Microsoft's announced retirement of Purview Data Governance adds migration planning to a week centered on verifying model behavior, infrastructure performance, and data access rather than assuming successful execution or platform compatibility.
This Week's Overview
- Azure compute targets data preparation, inference, and cluster reliability
- Model backdoors challenge behavioral evaluation
- Fabric Warehouse changes configuration and deployment workflows
- OneLake storage for Databricks reaches Fabric general availability
- Governance migration must preserve context and account for gaps
- Windows local inference expands APIs and optimization options
- Agent reliability needs checks beyond a successful tool call
- Fabric medallion pipelines need explicit data and access boundaries
- Other Machine Learning News
- Also relevant this week
- Dates to know
Azure compute targets data preparation, inference, and cluster reliability
Azure's compute updates cover dense CPU workloads, GPU inference benchmarks, and rack-scale validation. The reports distinguish preview performance estimates from measured serving results and operational testing.
HDv2 preview for data preparation and agent sandboxes
Azure HDv2 preview registration is open for AI data preparation, compute-intensive high-performance computing (HPC), and dense agentic sandbox workloads. Each VM uses AMD EPYC 9006 processors with more than 500 physical Zen 6 cores, 3.8 TB of DDR5-8800 memory, 30 TB of PCIe Gen 5 NVMe storage, and 400 Gb Azure Boost Networking. That combination targets workloads that need many parallel execution environments alongside large working datasets.
Microsoft reports approximately twice as many parallel sandboxes and up to 1.7 times more agentic task throughput than the next-best in-market alternative. These comparisons are preliminary, and the throughput estimate uses SPEC CPU 2026 Integer Rate as a proxy rather than direct measurements of agent tasks. Teams evaluating the preview should therefore test their own sandbox density and task mix before using those figures for capacity planning.
GPU inference comparisons
ND GB200 v6 testing covered 264 matched comparisons against bare metal, using four GPUs per node and workloads including DeepSeek-R1-0528, GPT-OSS-120B, and Kimi-K2.5. ND GB300 v6 testing covered 91 matched DeepSeek-R1-0528 configurations against the published InferenceX reference, with throughput, decode speed, and end-to-end latency closely matching that reference across tested serving stacks and precisions. The GB200 results included workload-dependent throughput and first-token gains, but differences between the VM and bare-metal environments mean those gains cannot be attributed to virtualization alone.
Validation for AMD Helios racks
Each Azure ND MI455X v7 compute rack contains 72 MI455X GPUs, 18 AMD EPYC Venice CPUs, and 31 TB of HBM4 memory, with a Microsoft-owned, two-tier Ethernet backend connecting racks into larger clusters. Microsoft tests firmware, operating-system images, drivers, and diagnostic tools together as a versioned baseline, then uses SuperBench for sustained burn-in workloads and RCCL AllReduce and AllGather tests to expose communication failures. Faulty nodes move to quarantine racks for diagnosis and repair, followed by workload validation before they return as healthy spares.
Model backdoors challenge behavioral evaluation
Two research reports examine ways attackers could alter a model before deployment. Both point to a gap between apparently normal task performance and confidence that a checkpoint is free of hidden behavior.
Poisoned post-training data can redirect coding agents
Researchers injected 90 poisoned examples into roughly 9,000 terminal-agent conversations, about 1% of the corpus, across 10 mock harnesses. Fine-tuned GPT-4.1 mini produced unsafe actions in 78.0% and 77.8% of trials on seen harnesses, while success on unseen harnesses fell to 11.0% and 0%. In a separate model-poisoning configuration, Qwen3.5-4B and Qwen3.5-9B reached 84.1%-94.8% backdoor activation on unseen repositories.
Reinforcement learning recovered much of the models' benign coding performance without materially reducing backdoor success. That makes ordinary coding evaluations an incomplete check for this kind of compromise. The findings require a backdoored checkpoint and do not establish weaknesses in default Claude Code or Codex CLI configurations.
Small weight edits can hide triggered commands
Another report shows that edits to fewer than ten neurons can create a sleeper agent that passes evaluations but emits an attacker's command when triggered, with implantation requiring only enough RAM to run the model. Microsoft has not observed threat actors weaponizing this technique. It recommends weight audits before deployment and after every update, alongside behavioral backdoor scanning, because Defender for Cloud model scanning addresses serialization and code-level risks rather than replacing weight-level audits.
Fabric Warehouse changes configuration and deployment workflows
Fabric Warehouse has an upcoming SQL pool compatibility change alongside a preview of expanded continuous integration and delivery (CI/CD). Client applications need migration attention, while schema deployment gains a version-controlled workflow.
SQL pool unification requires client changes
Microsoft Fabric Data Warehouse will unify built-in and custom SQL pool configurations beginning January 2027. Existing workspaces retain active configurations and workload behavior, but inactive configurations will disappear from the interface and GET responses. The customSQLPoolsEnabled flag will also be removed from PATCH requests, requiring dependent applications to update their client logic.
Customers should stop relying on customSQLPoolsEnabled to retrieve inactive configurations by mid-December 2026. New workspaces will receive a single read-optimized SQL pool, allowing SELECT and non-SELECT workloads access to up to 100% of available compute. For existing integrations, the immediate work is checking configuration retrieval and update code rather than assuming workload behavior will change.
Warehouse definitions enter Git and deployment pipelines
In preview, Fabric Warehouse Git integration represents a Warehouse as a SQL database project, supporting branches, commits, and pull-request review. DacFx compares source and target database states to generate incremental schema deployments, with versioned pre- and post-deployment scripts executing automatically around them. Warehouse Item Definition APIs support SQL project and DACPAC formats for creating, retrieving, and updating definitions, while deployment configurations, Warehouse support in fabric-cicd, and CI/CD for the SQL analytics endpoint remain planned capabilities.
OneLake storage for Databricks reaches Fabric general availability
Following last week's coverage of native storage for Azure Databricks, OneLake external-location support for Unity Catalog is generally available in Microsoft Fabric, while the Azure Databricks side remains in preview. Supported Unity Catalog-managed tables, volumes, views, materialized views, and streaming tables can use OneLake as their underlying storage. Fabric users access selected catalogs, schemas, and tables through mirrored Azure Databricks catalog items without copying the data, while Unity Catalog controls data access.
Setup requires a capacity-backed Fabric workspace, an Azure Databricks Storage item, and an approved managed identity or service principal. OneLake Security is enabled by default, no user receives access when the storage item is created, and workspace roles do not grant access to its data. The approved Azure Databricks identity must be the only member of the item's ReadWrite OneLake security role, making identity configuration a prerequisite rather than an optional hardening step.
- Use OneLake as native storage for Azure Databricks
- Use OneLake as native storage for Azure Databricks (Preview)
Governance migration must preserve context and account for gaps
The retirement of Purview Data Governance creates both a migration deadline and a need to assess replacement capabilities. The accompanying guidance separates reusable business context from the catalog that currently stores it.
Preserve business meaning before choosing a replacement
Microsoft announced the retirement of Purview Data Governance on November 12, 2026, with support ending on November 12, 2027. The retirement analysis identifies OneLake Catalog as the center of Microsoft's data governance, but recommends inventorying business definitions, ownership, lineage, rules, classifications, and relationships before selecting replacement technology. The concern is preserving the meaning attached to data, not merely moving catalog entries.
Fabric IQ's Ontology feature binds business concepts and relationships to data for AI consumption. The authors recommend governing reusable business terms outside individual data platforms and serving that context through Fabric ontologies. This approach treats the migration as a review of business knowledge and ownership rather than a direct catalog export-and-import exercise.
Check OneLake Catalog's current policy coverage
OneLake Catalog organizes discovery and governance into Explore, Govern, and Secure experiences, but Fabric capabilities do not provide one-to-one replacements for external metadata scanning, glossary depth, data quality workflows, or broad policy coverage. Fabric Policies currently cover external data sharing, item creation, and workspace settings editing, with only an Allow effect, limits of 100 rules at tenant scope and 50 at capacity scope, and activation taking about 15 minutes. Purview provides sensitivity labels, data loss prevention (DLP), audit, insider risk, and AI protection, with additional licensing required for Information Protection and DLP.
Windows local inference expands APIs and optimization options
Windows AI development now spans experimental task-specific runtime APIs, a local coding model checkpoint, and work on hardware-aware compilers. The reports address different layers of local inference, from application integration to memory use and execution scheduling.
Windows ML previews task-specific runtime APIs
Windows ML adds experimental llama.cpp support for running GGUF models locally through task-specific APIs. The experimental Windows ML Runtime API is in preview, with Text Generation and Speech Recognition APIs in the first release and an OpenAI-compatible endpoint for local prototyping. Existing ONNX Runtime APIs remain fully supported.
The runtime supports zero-copy Windows-native inputs, deterministic multi-model pipelines, and ahead-of-time model loading and compilation. Examples connect Triton kernel compilation with PyTorch, ONNX model export, and model preparation and benchmarking through Windows ML CLI. Developers can evaluate the new task APIs without treating them as a replacement requirement for existing ONNX Runtime integrations.
Local coding inference trades memory against throughput
MAI-Code-1.1-Flash is a 137.9-billion-parameter mixture-of-experts model with a released 53.0 GiB mixed-precision checkpoint, available through Microsoft Foundry and as the local model in GitHub Copilot. At 120K input tokens on RTX Spark, speculative decoding increased median throughput from 25.81 to 51.12 tokens per second, while sliding-window attention reduced reserved draft attention-cache memory at 262K context from 5,120 MiB to 1,156 MiB without retraining. These measurements used 128 GB unified memory and a 64 GB GPU allocation, so they should not be generalized to other serving conditions.
Compiler performance depends on data movement
Microsoft's compiler analysis explains why layer-by-layer execution may achieve only 5%-10% of expected throughput when memory access constrains computation. It argues for progressive lowering through multiple intermediate representations, hardware-aware tiling, scheduling, and fusion across heavyweight operations such as matrix multiplication and convolution. For client inference that may require batch size 1 and latency around 2 milliseconds, Microsoft is exploring DirectX Compute Graph Compiler, Windows ML, and multi-level intermediate-representation architecture to reduce execution inconsistencies across hardware.
Agent reliability needs checks beyond a successful tool call
Agent reliability guidance this week focuses on verifying outcomes rather than trusting completion messages. Architecture, training, and multilingual query evaluation each expose a different way an apparently successful run can fail.
Keep authority in trusted services
An illustrative Azure architecture separates experience and reasoning, trusted action, and operational intelligence into three logical planes. Models propose actions, while trusted services enforce authorization, policy, and authoritative state before changing records, with Fabric handling governed outcome analytics outside booking, payment, and command execution. Unknown partner execution outcomes require reconciliation before retries because duplicate detection alone cannot resolve uncertain bookings, and the proposed pilot path moves from shadow mode to assisted booking and bounded autonomy rather than claiming a verified production reference architecture.
Train against failures that appear complete
The expanded ThinkingBox study evaluates 18 models across 507 benchmark tasks with 20 attempts per task, finding that 67.24% of 79,853 failed trials appeared complete according to final responses, state-changing tool calls, and error-free final tool responses. After 50 reinforcement learning steps, Qwen3.8-27B improved single-attempt success from 51.70% to 60.78%, while tasks succeeding on all 20 attempts increased from 38 to 107. Estimated cost per all-20 task fell from $24.36 to $8.00, but that estimate excludes training, task-selection probes, simulated users, judges, and infrastructure.
Normalize multilingual filter values and retest after migration
An exploratory test using one synthetic 288-row Fabric warehouse table graded 42 responses and found that all 10 incorrect answers used Traditional-script filter values absent from the table's stored Simplified Chinese values. A Foundry normalization route produced seven correct answers and three incomplete runs, with no incorrect completed answers. Microsoft Learn lists November 6, 2026 as the Fabric data agent tool's deprecation date and directs users to Fabric IQ preview, which the study did not test, so the article recommends repeating independent SQL-based evaluations after migration.
Fabric medallion pipelines need explicit data and access boundaries
Fabric guidance covers both constructing layered data pipelines and securing them. Separating staging, validated records, and aggregates does not by itself establish access boundaries.
Build traceable record versions from source files
A medallion architecture example uses Dataflow Gen2 to discover Excel workbooks in SharePoint and retain source filenames and modified timestamps in staging. The Bronze notebook uses SHA-256 hashes of business attributes and a left anti join to identify unseen record versions, while Silver validates required columns, standardizes values, and selects the latest observation per customer. Gold aggregates customer counts, active customers, annual revenue, regions, and segments, but vendor-spend analytics is outside the demonstrated implementation.
Secure SQL access separately from underlying files
Fabric Warehouse ReadData and ReadAll are separate permissions, and warehouse row-level and column-level security do not protect underlying files accessed through Spark. Workspace Admin, Member, and Contributor roles grant CONTROL on warehouses and SQL analytics endpoints, including unmasked access, while Power BI Direct Lake models fall back to DirectQuery to honor warehouse row-level and column-level security. Dynamic data masking does not prevent inference through range filters, so the guidance recommends separate medallion workspaces for access boundaries or granular SQL permissions when layers share a warehouse.
Other Machine Learning News
mssql-python now supports opt-in SQL Server result fetching as Apache Arrow structures through three cursor APIs for individual RecordBatch fetching, eager Table fetching, and lazy RecordBatchReader streaming, without changing fetchone, fetchmany, or fetchall. Tests found faster fetching for most SQL Server types, including benefits from C++ temporal conversions, but NVARCHAR fetching is currently slower on Linux and awaits a follow-up fix.
Azure Monitor's global Grafana Model Context Protocol (MCP) endpoint requires no Azure Managed Grafana resource or server deployment, providing 21 tools under the signed-in user's existing Microsoft Entra ID permissions. Twenty tools are read-only and one updates saved dashboards with validation and version checks, but the endpoint supports user sign-in only and cannot create dashboards or access Grafana alerting, Loki, SQL, or ClickHouse tools.
Azure NetApp Files guidance covers checkpoint writes, high-throughput NFS reads, quality-of-service controls, writable short-term clones, and Trident-provisioned persistent volumes for inference on Azure Kubernetes Service. Its object REST API exposes NFS and SMB datasets to Microsoft AI and data services without a separate object-store copy, although the reference deployments report qualitative improvements rather than numerical performance comparisons.
- Azure NetApp Files: The Data Backbone for NVIDIA AI Model Training, Inference, and Microsoft Foundry
A practical agent-skills guide defines a skill as a versioned directory centered on SKILL.md, with the host responsible for tools, credentials, and permissions, and demonstrates shared instructions packaged for Claude Code, Cursor, and GitHub Copilot. Its Data Forge support-desk workflow has seven phases with human checkpoints, while Proof Agent returns evidence-backed PASS, FAIL, or PARTIAL results from static diff review without executing code or tests.
The October Fabric community spotlight collects 46 posts, videos, and tutorials covering topics including Fabric IQ, data-agent and semantic-model governance, deployment workflows, Spark performance, and Power BI modeling. It also includes FabCon recaps and a free DP-700 study course.
Also relevant this week
- What’s new in AKS networking: Connectivity, performance, and observability: Cross-cluster networking for Azure Kubernetes Fleet Manager is generally available, supporting service discovery, communication, network policy enforcement, and observability across AKS clusters. Full story in the Azure roundup.
- What's new in FinOps toolkit 15 – October 2026: FinOps toolkit 15 adds an agent plugin with five agents and a read-only Azure MCP server scoped to Data Explorer. Full story in the Azure roundup.
- From Python Agents to Copilot Harness : Putting Microsoft Foundry to the Test: The experiment replaces four hosted Microsoft Agent Framework wrappers with one Foundry-managed Copilot harness agent. Full story in the Artificial Intelligence roundup.
- What’s new across Microsoft SQL in 2026 so far: the PASS Summit West roundup: Hyperscale elastic pools preview supports up to 50 databases and 128 TB per pool, depending on hardware and service objective. Full story in the Azure roundup.
Dates to know
Deadlines
- December 2026: Microsoft Fabric SQL pools: API clients must stop relying on customSQLPoolsEnabled to retrieve inactive configurations by mid-December 2026. (Unifying SQL Pools: Upcoming changes to UI and APIs)
- January 2027: Microsoft Fabric SQL pools: January rollout unifies SQL pool configuration and removes customSQLPoolsEnabled from PATCH requests; existing active configurations remain unchanged. (Unifying SQL Pools: Upcoming changes to UI and APIs)
- 12 November 2027: Microsoft Purview Data Governance: Support for Purview Data Governance ends. (Purview Data Governance is retiring: Your real asset isn’t the catalog, it’s the context)