Weekly Security Roundup: Git Fixes and AI Trust Boundaries

Git 2.56.1 fixes six vulnerabilities, with backports and patched GitHub Enterprise Server releases available, but developers must still update bundled Git copies even though GitHub's cloud services are patched. November .NET servicing addresses multiple vulnerability classes, while Windows tightens PowerShell signature verification and PowerShell 7.7 strengthens execution defaults, making compatibility testing part of patch deployment. GitHub added opt-in workflow-log secret scanning and repository-level AI Scan API controls, and announced that Dependabot alerts will pause on most repositories inactive for one year starting December 8 unless policy exemptions apply. Research into poisoned coding-agent models and small weight-level backdoors shows why restored coding performance and ordinary evaluations do not establish model safety. Across agent tooling, Kubernetes policy, and remote-access investigations, the practical message is to verify what controls actually enforce: connectivity is not authorization, auditing is not blocking, and application patching does not remove host persistence.

This Week's Overview

Security servicing spans Git, application runtimes, and Windows

The week's patching priorities include Git vulnerabilities, .NET servicing, and Windows security changes that can affect script execution. The OpenJDK articles cover earlier servicing releases, so their release dates and version numbers need to remain distinct from November's updates.

Git fixes require checking bundled copies

Six Git vulnerabilities can cause code execution, path traversal, memory corruption, crashes, or unauthorized filesystem access. Fixes ship in Git 2.56.1, with backports to 2.55.1, 2.54.1, 2.53.1, 2.52.1, 2.51.3, and 2.50.2. Patched GitHub Enterprise Server versions are 3.19.15, 3.20.11, 3.21.9, and 3.22.4.

GitHub has deployed fixes to GitHub.com and GitHub Enterprise Cloud, including deployments with data residency. Developers still need to update bundled Git copies and avoid untrusted clones, merges, binary patches, and fast-import streams until those copies are patched.

.NET servicing covers several vulnerability classes

The November 10, 2026 servicing releases for .NET 10, .NET 9, and .NET 8 include security and non-security fixes. Release resources cover .NET 11.0.0, .NET 10.0.15, .NET 9.0.23, and .NET 8.0.34. Listed advisories address elevation of privilege, denial of service, information disclosure, security feature bypass, and Visual Studio remote code execution vulnerabilities.

The release table is not a complete affected-product inventory, so teams should use each advisory's full product list and remediation guidance when planning updates. Deployment resources include installers, binaries, Linux package instructions, and container images through Microsoft Container Registry.

Windows changes signing behavior and PowerShell verification

Released November 10, 2026, KB5132061 applies to Windows 11 versions 26H2, 25H2, and 24H2. The November release enables clients to offer NTLM message signing by default for non-anonymous connections and adds auditing capabilities. Stricter PowerShell signature verification can reject scripts with invalid UTF-8 encoding, so affected scripts should use valid UTF-8 with a byte order mark and be re-signed.

Hotpatch-enabled devices need separate PowerShell security packages after the November hotpatch update: KB5135385 for applicable Windows 11 versions and KB5135386 for Windows Server 2025. Microsoft will skip the December 2026 non-security preview update during the month's second half, with normal monthly servicing resuming in January 2027.

PowerShell 7.7 tightens execution defaults

PowerShell 7.7 is generally available as a Standard Term Support release built on .NET 11. Security changes strengthen remoting validation and restrict property assignment on non-core objects in Constrained Language Mode. Test-Json no longer resolves external schema references by default, requiring -FetchExternalReference when scripts need that behavior.

Compatibility testing should cover non-security changes as well. New-Guid now generates version 7, time-sortable UUIDs by default instead of version 4 UUIDs, while New-TemporaryDirectory creates a uniquely named directory and returns a DirectoryInfo object.

OpenJDK releases clarify patch levels and support planning

The October 2026 Microsoft Build of OpenJDK patch release includes 25.0.5.0.2, 21.0.13.0.2, 17.0.21.0.2, and 11.0.33, with no Microsoft-specific updates for those JDKs. Its vulnerability advisories are dated October 20, 21, and 24, 2026. Support for Microsoft Build of OpenJDK 11 and 17 ends in September 2027, making runtime migration planning relevant alongside patch deployment.

The earlier July release provided 25.0.4, 21.0.12, 17.0.20, and 11.0.32, updating processor enumeration across processor groups and marking Reserved Stack Area unsupported on Windows/ARM64. August provided 25.0.4.1, 21.0.12.1, 17.0.20.1, and 11.0.32.1, adding G1 time-based heap uncommit in OpenJDK 25 through the diagnostic G1UseTimeBasedHeapSizing flag, enabled by default. Microsoft continues supporting Eclipse Temurin OpenJDK 8 on Azure and providing container images based on Azure Linux and Ubuntu.

.NET strengthens memory safety and expands cryptography

Beyond servicing existing releases, the .NET sessions examine changes to runtime responsibilities and cryptographic capabilities. These are separate development concerns: auditing unsafe code and preparing applications for future cryptographic requirements.

Runtime changes make unsafe code easier to audit

The updated .NET memory safety model introduces stricter compile-time enforcement and improved auditability of unsafe code use. That gives developers a more explicit basis for reviewing code that operates outside memory-safety guarantees.

Runtime Async moves asynchronous orchestration and code generation from the Roslyn compiler into RyuJIT in the .NET Runtime. The move is intended to improve performance, while the memory-safety changes address enforcement and reviewability rather than making a performance claim.

Post-quantum capabilities inform migration preparation

.NET 11 expands the platform's post-quantum cryptography capabilities. The session covers both present-day use and preparation for future cryptographic requirements, providing a planning resource without specifying particular algorithms or deployment requirements in the supplied details.

Repository and dependency protection depends on explicit controls

Repository hardening guidance distinguishes what immutable releases, workflow restrictions, and dependency controls actually protect. Several controls have narrower coverage than their names might suggest, particularly around movable tags, installation-time verification, and publishing permissions.

GitHub Action repositories face a default event restriction

GitHub Actions will block pull_request_target events by default in public repositories from December 2, 2026, unless an applicable event policy explicitly allows them. Maintainers should review those policies separately from workflow actor policies, which restrict who can trigger matching workflows but do not control job-token permissions or publishing through other workflows and APIs.

Immutable releases protect published tags and assets, but not release notes, and cover only releases published after immutability is enabled. Floating major and minor tags must remain separate when maintainers intend to move them. Workflow dependency lockfiles record action commits and repository identities, including supported transitive dependencies, but do not pin job-level reusable workflows.

CLI extension publishing is not installation verification

GitHub CLI extension installation does not automatically verify release assets or attestations. Maintainers using immutable releases must upload assets and build provenance before publishing, and the protection applies only to releases published after immutability is enabled.

The guide identifies two further boundaries: GitHub Actions lock files do not pin container images, and installing an already-installed extension with --pin --force currently ignores the pin. Those limitations make it important to test the actual installation path rather than assume repository settings guarantee consumer-side verification.

Malicious dependencies can execute before application code

Compromised maintainer accounts and lookalike package names can introduce malware that executes during installation, before an application calls the dependency. GitHub's package firewall blocks known malicious packages before installation, while Dependabot alerts developers when malware is already present in their dependencies.

For .NET teams, Jon Douglas's NuGet session addresses software supply chain security in the context of AI-assisted development. Together, the sessions distinguish prevention before installation from detection after a dependency has entered a project.

GitHub scanning changes coverage, activation, and alert handling

GitHub's scanning updates add detection surfaces and management options, but several require explicit enablement. A separate Dependabot change will pause alerts on most inactive repositories unless policy-based exemptions apply.

Dependabot will pause alerts after repository inactivity

Starting December 8, 2026, GitHub will pause Dependabot alerts on most repositories inactive for one year and archive existing alerts. Repository creation, pushes to any branch or tag, and enabling alerts count as activity. Any push or manually enabling alerts restores archived alerts and restarts the inactivity clock.

Enforced security configurations and enterprise policies that prevent repository administrators from toggling alerts exempt repositories from these pauses. The change covers all github.com plans and GitHub Enterprise Cloud with data residency, but not GitHub Enterprise Server. Teams retaining deployed code in quiet repositories should check whether those repositories have the intended policy protection.

AI Scan gains repository-level API management

The AI Scan REST API public preview adds repository_scan to organization and repository endpoints. Organizations can disable repository scanning centrally or enable inheritance with repository opt-outs, and updates to pr_scan and repository_scan are atomic: if either change is blocked, neither applies and the request returns HTTP 422.

Enabling repository AI Scan does not run a scan. Users with write access start scans on demand, consuming GitHub Code Security AI credits. The preview supports organization-owned public repositories and eligible private or internal repositories, but not GitHub Enterprise Server.

Workflow log scanning starts only after enablement

Secret scanning for GitHub Actions workflow logs is in public preview for GitHub Enterprise Cloud enterprises with GitHub Secret Protection or GitHub Advanced Security. Enterprise owners must enable the disabled-by-default policy, and only eligible repositories with secret scanning enabled are covered. Runs completed before enablement are not backfilled.

The preview detects provider patterns, excluding generic patterns, custom patterns, AI-detected secrets, and secrets already masked by Actions. Findings appear in repository secret scanning alerts without notifications during preview and remain available after the workflow logs are deleted.

Cloudflare service tokens gain a detector

GitHub secret scanning now recognizes Cloudflare service tokens through the cloudflare_access_service_token secret type. Partner secrets found in public repositories are reported to their issuer, while user secrets found in public or private repositories generate secret scanning alerts.

Public monitoring alerts become directly shareable

Individual public monitoring alerts now have direct links that authorized teammates can bookmark or share. Clicking an alert number in the enterprise Public monitoring view opens the alert, and the feature remains available to GitHub Enterprise Cloud customers with GitHub Secret Protection or GitHub Advanced Security at no additional cost.

Memory-safety research exposes validation and remediation costs

Linux and Rust investigations show different sides of vulnerability research: turning automated findings into upstream fixes, and checking safety assumptions around panics and external state. Neither investigation supports treating every finding as equally exploitable.

Linux findings require substantial work after scanning

MDASH scanning generated 26,502 raw findings, narrowed to 621 proof-of-concept-backed cases and 59 separate upstream reports. Those reports received 33 assigned vulnerability identifiers, including 24 rated High or above under Common Vulnerability Scoring System v3.1. Fixes for 38 reports had merged upstream, with a median 41 days from the first public patch to mainline release.

CVE-2026-64563 enabled stable local privilege escalation on Linux and affected Android with SELinux enforcing, without special kernel configurations or user namespaces. Reporting the 59 validated issues required more than four person-months after approximately two weeks of scanning. The results underline that discovery throughput and the work required to validate, report, and land fixes are different measures.

Rust findings examine panic and operating-system assumptions

FORGE Lab investigated Rust standard-library soundness bugs involving panic safety, user-provided implementations, and operating-system state. Some findings came from Hyenas, FORGE's frontier research version of the AI-powered vulnerability-scanning system codename MDASH. Issues affecting strip_prefix_of and split_once remained open on November 9, 2026.

Rust 1.100, scheduled for November 12, requires Allocator implementors never to unwind from trait methods or drop. FORGE assessed practical exploitation of the signal-stack cleanup bug as unlikely under tested conditions, but did not extend that assessment to the other findings. Developers reviewing these reports should preserve that distinction rather than generalize one exploitability result.

Remote-access investigations require host-level remediation

The ScreenConnect and GSocket reports both show why closing an entry path does not necessarily remove an attacker's continued access. Their remediation guidance extends beyond application patching to persistence and exposed sessions.

SCKit targets endpoints and technician sessions

Microsoft confirmed reported ScreenConnect abuse and observed unauthorized remote-access clients and SCKit on customer-managed endpoints. The initial entry method remains under investigation, and reported RMM+ extension abuse is separate from RelayTrap's endpoint-to-technician attack path.

ScreenConnect client 26.6.5 and later block the attempted endpoint-to-technician transfer technique, but technician clients must be updated separately from the console. Removing unauthorized clients alone does not eliminate SCKit persistence or invalidate exposed authenticated web sessions. Microsoft provides Defender detections and hunting queries for infrastructure, command lines, file hashes, registry persistence, scheduled tasks, and antivirus exclusions.

GSocket persists beyond application patching

Devices with GSocket-related detections increased 356.2% from March through August 2026 against the March baseline, and web applications and hosting accounted for 87.1% of attributable incidents in the analyzed Microsoft Defender for Endpoint telemetry. GSocket uses shared public relay infrastructure and outbound connections, allowing attackers to reconnect without knowing the compromised host's IP address.

Application patching does not remove the GSocket client or its persistence mechanisms, so remediation must address the host itself. Some incidents included attempts to exploit CVE-2026-43284 for Linux privilege escalation, but the analysis did not establish successful escalation in every case.

Model backdoors complicate AI supply chain validation

Two research reports examine malicious behavior embedded in models rather than ordinary application code. Their findings make model provenance and weight-level assessment relevant alongside behavioral testing, while retaining important limits on what the experiments demonstrate.

Poisoned post-training data can preserve hidden behavior

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, compared with 11.0% and 0% on unseen harnesses. The strongest model-poisoning configuration achieved 84.1% to 94.8% backdoor activation on unseen repositories using Qwen3.5-4B and Qwen3.5-9B.

Reinforcement learning recovered much of the models' benign coding performance while leaving backdoor attack success largely unchanged. That means restored coding performance did not establish that the tested backdoors had been removed. The harness-transfer results do not indicate weaknesses in default Claude Code or Codex CLI configurations: the risk requires a backdoored checkpoint.

Small weight edits can evade ordinary evaluations

Edits to fewer than ten neurons can create a sleeper agent that passes evaluations and emits an attacker's command when triggered. Implanting the weight-space backdoor requires only enough RAM to run the model, although Microsoft has not observed threat actors weaponizing this technique.

Microsoft recommends auditing model weights before deployment and after every update, alongside behavioral backdoor scanning. Defender for Cloud model scanning addresses serialization and code-level risks, but does not replace weight-level audits.

Kubernetes controls distinguish visibility from enforcement

Azure Container Linux and AKS updates add integrity auditing, cross-cluster policy support, and more selective enforcement. Across these changes, recording a violation is not the same as blocking it.

Azure Container Linux adds inventory and integrity controls

Azure Container Linux now supports AKS node auto-provisioning, selecting virtual machine sizes and provisioning nodes according to workload requirements. Confidential VM support covers Intel TDX and AMD SEV-SNP, with FIPS-enabled confidential VM images also being introduced. The CIS Benchmark for Azure Container Linux is now generally available.

Optional IPE audit mode records execution attempts that violate integrity policies without blocking them, giving teams a way to inspect violations before enforcement. Each image now embeds an SPDX 2.2 JSON software bill of materials, but vulnerability assessment requires scanners that integrate with that inventory. The embedded inventory alone does not perform the assessment.

AKS networking extends policy across clusters

Cross-cluster networking for Azure Kubernetes Fleet Manager is generally available, supporting service discovery, communication, network policy enforcement, and observability across AKS clusters. Multi-NIC pods are in public preview, allowing separate network interfaces for application and storage traffic. Managed DRANET support for remote direct memory access network interfaces is generally available, with topology-aware allocation for distributed AI and high-performance computing workloads.

Configurable observability dimensions are in public preview. They allow targeted collection of network data for troubleshooting, security investigations, and compliance, making data-collection configuration part of the operational security design.

Azure Policy separates admission blocking from auditing

Azure Policy supports separate enforcement points for Gatekeeper auditing, Gatekeeper admission validation, and Validating Admission Policy. A deny effect scoped only to auditing reports compliance without blocking violating resources at admission. Validating Admission Policy runs before Gatekeeper validation, with Gatekeeper providing a fallback when both are selected.

Each assignment can select multiple enforcement points but only one enforcement action. The article's JSON examples reference undeclared parameters, so they are illustrative rather than a complete deployment-ready assignment and definition pair.

Cosmos DB connects key rotation with scoped access

Azure Cosmos DB's key-usage visibility helps teams prepare credential changes, while its access-control guidance explains why scoped roles and account keys need separate treatment. Moving to narrower permissions requires checking every dependency that still uses key-based authentication.

Safe Key Rotation becomes generally available

Safe Key Rotation is now generally available and can be enabled directly from the Azure portal's Features blade. Azure Cosmos DB provides the last usage timestamp for each account key, allowing customers to check usage before regenerating keys or disabling local authentication.

The practical value is evidence for migration and rotation decisions rather than an assumption that an old key is unused. Teams can use the timestamps to check whether dependencies still rely on a key before changing it.

Control-plane roles do not grant item access

Azure Cosmos DB for NoSQL uses Azure role-based access control for account configuration and native data-plane roles for item access. Azure Owner and Contributor roles do not automatically grant item access, and data roles do not grant account configuration permissions. Data-plane roles can be scoped to an account, database, or container.

The built-in Data Contributor role includes delete permission, so applications that must write without deleting need a custom role. Account keys can bypass scoped data permissions while key-based authentication remains enabled, and local authentication should be disabled only after all migrated dependencies have been tested.

Entra External ID expands native authentication with explicit boundaries

Microsoft Entra External ID has generally available additions for social sign-in, second-factor authentication, watch companion apps, and embedded web views. Each feature has a distinct boundary around token issuance, browser delegation, or application responsibility.

Social sign-in retains browser-delegated steps

Social identity provider support for native authentication is generally available for native and single-page applications. Google, Facebook, and Apple authentication uses browser-delegated flows to comply with provider OAuth requirements. Subsequent steps, including multifactor authentication when Conditional Access is enabled, remain in that browser-delegated flow.

Microsoft Entra External ID issues ID and access tokens only after successful social provider authentication. Native experiences for applicable post-social authentication steps are planned for a future release, not part of the current capability.

Email and SMS become native second factors

Email and SMS one-time passwords are generally available as second-factor multifactor authentication for native authentication. Microsoft Entra Conditional Access evaluates and enforces the second factor after first-factor authentication succeeds. ID and access tokens are issued only after all required factors succeed, without client-side enforcement logic.

Developers configure Conditional Access policies in the External ID tenant and integrate through native authentication SDKs or APIs. The enforcement responsibility remains with the identity service rather than application code deciding whether the additional factor can be skipped.

Apple Watch token transfer requires lifecycle handling

Refresh token transfer is generally available, allowing Apple Watch companion apps to refresh access tokens without relying on phone connectivity. Refresh token access requires explicit opt-in application configuration, and applications can securely transfer tokens from iOS using WatchConnectivity.

Developers remain responsible for token rotation, revocation, secure storage, and secure transport. Independent access from the watch therefore adds a token-management responsibility rather than removing one.

Native-to-web single sign-on stays within one application

Native authentication now supports generally available single sign-on from native apps to embedded web views such as WKWebView and Android WebView. The native app obtains an access token and loads the web view with an Authorization: Bearer header. The web resource must validate that token before granting access without reauthentication.

This flow does not rely on browser session state. The release supports single sign-on within one application, while cross-app and cross-device scenarios remain areas of investment.

Entra application design accounts for replication and policy reuse

Entra guidance addresses two separate implementation concerns: asynchronous directory replication and repeated Conditional Access configuration. Both require understanding the identity platform's behavior rather than relying on client-side assumptions.

Successful writes do not guarantee immediate reads

Microsoft Entra accepts directory writes before asynchronously replicating them, so a newly created object can temporarily return Not Found after a successful write. This applies to Microsoft Graph applications and higher-level tools such as PowerShell. Application-only access does not guarantee consistency across requests, while delegated access may provide more predictable consistency.

Recommended patterns include trusting successful writes, caching returned identifiers, and minimizing read-after-write dependencies. Retries should use exponential backoff and idempotency rather than treating every temporary missing-object response as a failed creation.

Conditional Access plans reduce repeated attributes

Conditional Access policy plans provide a way to avoid repeating common attributes across policies. John Savill's overview compares regular policies with plans, explains rules within plans, and covers migration from existing policies.

Agent tool access needs controls beyond connectivity

Database and infrastructure assistants expose operational tools through different permission models, including read-only defaults, delegated user access, and explicit write flags. The common implementation question is which identity authorizes each operation, not merely whether an agent can reach the service.

DocumentDB defaults to read tools

The open-source Azure DocumentDB MCP Toolkit is in public preview, exposing database operations to AI agents through Model Context Protocol. Read tools are enabled by default, while write and management tools require explicit environment flags. Administrator-defined connection profiles support Microsoft Entra ID authentication, hostname allow lists, and SCRAM-SHA-256 for local or non-Entra setups.

The toolkit supports Claude Code, Cursor, Codex, Gemini CLI, and GitHub Copilot, and can run directly through npx. Vector search and BM25 full-text search use the aggregate tool. Teams evaluating it should review both the enabled tool set and the connection profile's authentication and destination restrictions.

Private PostgreSQL still needs application authorization

Private database connectivity does not establish an application's identity, authority, or intent. The PostgreSQL guidance compares public MCP, private MCP, and public gateway architectures backed by private Azure Database for PostgreSQL Flexible Server. Azure App Service virtual network integration provides outbound connectivity, not private inbound access.

Gateway-backed designs must prevent alternative public access to the MCP backend that could bypass Azure API Management policies. Restricted workload permissions and database row-level security then constrain data access independently of MCP tool authorization.

SQL MCP auto configuration exposes permissive defaults

Data API builder 2.0 introduced auto configuration in June 2026, using autoentities wildcard patterns to include SQL tables. The bundled container configuration exposes every table in every schema, enables development tooling, and allows anonymous create, read, update, and delete operations. Those defaults are unsuitable for production without review.

DAB_CONNSTRING overrides the primary configuration's connection string, including existing values or function references. Deployment examples use image version 2.0.12, making configuration review necessary even when the container starts successfully with minimal setup.

FinOps adds a scoped assistant and private-deployment options

FinOps toolkit 15 adds an agent plugin with five agents and a read-only Azure MCP server scoped to Data Explorer. FinOps hubs 15 introduces optional NAT Gateway support for private deployments alongside virtual network and public networking modes. Local hubs require Docker with at least 16 GB of container memory, PowerShell 7, and FOCUS cost exports in Parquet format.

Operational changes need attention independently of agent permissions. Corrected commitment discount eligibility columns require historical Prices data to be reingested, and Azure Optimization Engine now uses a Data Collection Rule and Ingestion API following the Data Collection API's September 2026 deprecation.

Database Hub Copilot preserves existing permissions

Copilot in Database Hub is in public preview, answering questions from resource inventory, detected issues, and performance telemetry. It respects existing Azure and Fabric permissions and does not change databases. A Fabric administrator must enable the required tenant settings before users can access it.

The preview requires Fabric capacity F2 or higher or Power BI Premium capacity P1 or higher, with token usage consuming capacity units. OpenAI-operated models are not currently available in government or sovereign clouds, so permissions alone do not establish deployment eligibility.

Grafana MCP uses the signed-in user's authority

Azure Monitor dashboards with Grafana MCP provides a global endpoint without requiring an Azure Managed Grafana resource, MCP server deployment, or Grafana data source configuration. Its 21 tools include 20 read-only tools and one that updates existing saved dashboards. Every call uses the signed-in user's Microsoft Entra ID token and existing permissions, and signing in grants no additional access.

Dashboard updates support validation without saving and version checks that prevent overwriting intervening edits. The endpoint supports user sign-in only and cannot create dashboards or access Grafana alerting, Loki, SQL, or ClickHouse tools.

Copilot Studio connection changes preserve organizational controls

Microsoft expects the improved Copilot Studio connection experience to reach general availability by December 31, 2026. Users will configure third-party connections within conversations, with eligible valid connections reusable across agents accessed through Microsoft Copilot, Microsoft Teams, and Web Chat.

Custom web interfaces using the Standard harness must open the new connection window and resume conversations after successful setup. Existing organizational controls, authentication requirements, and user permissions will continue to govern access.

Agent execution separates model proposals from trusted authority

Agent architecture guidance focuses on keeping authorization, state changes, credentials, and execution isolation outside the model's discretion. The examples range from illustrative designs to migration reports and previews, so they should not be treated as interchangeable production reference architectures.

Trusted services own state changes and reconciliation

The accountable-agent architecture separates experience and reasoning, trusted action, and operational intelligence into three logical planes. Models propose actions, while trusted services enforce authorization, policy checks, and authoritative state before changing records. Unknown partner execution outcomes require reconciliation before retries, because duplicate detection alone does not resolve an uncertain booking.

The scheduler implementation follows a related boundary: agents propose placements, while application code owns calendar state and validates demand, duration, spacing, conflicts, and resource capacity. Its gateway returns 202 Accepted after persisting commands, with PostgreSQL queues and checkpoints supporting the workflow. These examples give concrete context to Niels Provos's warning about unrestricted agent access to workplace tools and hidden malicious behavior in third-party skills, while the accountable-agent pattern remains illustrative rather than verified for production.

Foundry delegation preserves downstream authorization

Foundry user delegation supports On-Behalf-Of access so downstream services can authorize requests using the user's delegated identity. Tool search retrieves relevant tools rather than loading every definition into each model request, while evaluation datasets can be generated from sampled traces or synthesized from instructions, prompts, and reference documents.

Insights, rubric evaluators, custom evaluators, and managed memory are described as preview capabilities. AI Gateway provides token-rate limits and total token quotas, not native dollar-budget enforcement in Agent Service. Adrian Twarog's introductory walkthrough complements that guidance by demonstrating an MCP server with scoped permissions and calling a published agent through the Azure SDK.

Harness experiments retain application-owned identity boundaries

One experiment replaces four hosted Microsoft Agent Framework wrappers with a single Foundry-managed Copilot harness agent. A React workbench and Node.js backend still handle authentication, streaming, conversation ownership, and authorized downloads. Fabric access uses the signed-in user's context, while the Foundry IQ connection uses the project's managed identity.

Early tests returned correct sales figures without retrieving the requested explanatory document, followed by one successful retrieval from both sources after instructions were clarified. The public-preview demo does not establish production readiness, feature parity, or improved quality, latency, or cost.

Sandboxes contain execution without correcting behavior

Patrick Nikoletich demonstrates a Dockerized agent worker pool with isolated containers and credential-safe inference gateways that avoid exposing credentials to workers. Agent Host Protocol tunnels CLI sessions into those containers, and the repository includes a harness builder and secure Aspire agent service. A Rust rewrite and direct Azure tunneling are Copilot Harness 2.0 plans, not demonstrated current capabilities.

A medical research assistant migration uses Foundry hosted agents with VM-isolated session sandboxes, while retaining telemetry and evaluation workflows. Its networking guidance recommends a /24 subnet for production and documents /27 as the supported minimum for smaller deployments. The Copilot CLI modernization session reinforces the underlying boundary: sandboxing can contain accidental deletion and unintended resource access, but does not eliminate the behavior that causes them.

Container deployment avoids long-lived pipeline credentials

The Azure Container Apps deployment guide connects images in Azure Container Registry to Container Apps and Azure OpenAI. Container Apps can reference Azure Key Vault secrets using managed identity, and the guide recommends workload identity federation using OpenID Connect for supported pipeline authentication instead of long-lived client secrets.

Multiple-revision mode supports traffic splitting, blue-green deployments, and canary releases. Scaling application replicas does not automatically increase Azure OpenAI model capacity, quota, or token throughput, so rollout and capacity planning remain separate tasks.

Model availability brings retention and approval requirements

Claude Fable 5.2 availability in GitHub Copilot and Microsoft Foundry comes with service-specific enablement and data-handling conditions. Administrators need to assess those conditions separately from whether a model appears in the product catalog.

GitHub Copilot requires administrator enablement

Claude Fable 5.2 is rolling out gradually to GitHub Copilot Pro+, Max, Business, and Enterprise, with provider list pricing under usage-based billing. Business and Enterprise administrators must enable its policy, which is off by default.

Anthropic retains prompts and outputs for safety classifiers but does not use retained data to train its models. Eligible enterprises can use Fable 5, 5.1, and 5.2 with zero data retention through 2026, while continued use afterward requires Enterprise Frontier Safeguards. Retention approval is therefore a separate decision from model enablement.

Foundry access distinguishes Fable from restricted Mythos availability

Claude Fable 5.2 and Fable 5.2 Fast are generally available in Microsoft Foundry, hosted on Anthropic with Global Standard and US DataZone deployment options. Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, while Fast doubles those prices and lacks batch pricing. Mythos 5.2 access is restricted to authorized customers through Anthropic verification programs, with eligibility confirmed through Microsoft account teams.

Blocked cybersecurity queries receive responses from Opus 4.8, while blocked biology queries receive responses from Opus 5. Both models require safety-monitoring data retention, although eligible customers can use Fable 5.2 with zero retention until Enterprise Frontier Safeguards becomes available. Deployment review should account for retention and response-model behavior, not only the selected model name.

Fabric governance must cover every data access path

Fabric guidance distinguishes storage permissions, warehouse SQL controls, and catalog governance. A restriction applied through one interface does not necessarily protect another route to the same underlying data.

OneLake and Databricks retain separate release and access boundaries

OneLake external-location support for Azure Databricks Unity Catalog is generally available in Microsoft Fabric, while the Azure Databricks side remains in preview. 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, and creating the storage item grants no user access.

Workspace roles do not grant access to the storage item's data, and the approved Azure Databricks identity must be its only ReadWrite OneLake security-role member. Supported Unity Catalog-managed tables, volumes, views, materialized views, and streaming tables can use OneLake as storage. Fabric users access selected data through mirrored Azure Databricks catalog items without copying it, while Unity Catalog controls data access.

Warehouse SQL restrictions do not protect direct file access

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 access to unmasked data. Direct Lake semantic models fall back to DirectQuery to honor warehouse row-level and column-level security.

Dynamic data masking reduces accidental exposure but does not stop users with query access from inferring masked values through range filters. Separate medallion workspaces provide access boundaries, while granular SQL permissions support layers sharing one warehouse.

Catalog governance is not a full Purview replacement

The OneLake catalog organizes discovery and governance into Explore, Govern, and Secure experiences. Fabric Policies currently cover external data sharing, item creation, and workspace settings editing, with only an Allow effect. Rules are capped at 100 for tenant scope and 50 for capacity scope, with activation taking about 15 minutes.

Purview supplies sensitivity labels, data loss prevention, audit, insider risk, and AI protection, with additional licensing required for Information Protection and data loss prevention. Fabric does not provide one-to-one replacements for external metadata scanning, glossary depth, data quality workflows, or broad policy coverage. The October community spotlight provides further guidance on securing and governing data agents and semantic models for AI among its 46 posts, videos, and tutorials.

Confidential VMs add disk protection and workload-specific benchmarks

Azure confidential computing updates address both protection of data disks and the performance of protected virtual machines. The benchmark reports describe particular configurations and workloads, not a universal confidentiality overhead.

Data disk encryption uses attested key release

Confidential Data Disk Encryption is in private preview for Azure confidential VMs. Encryption runs inside the confidential guest using LUKS2 on Linux or BitLocker on Windows, and Secure Key Release checks authorization and attestation claims before releasing key material.

Data disks need a separate Disk Encryption Set from confidential operating-system disks because their key policies differ. Customers must confirm subscription and region enablement with their Microsoft account team before evaluating the preview.

AMD results show gains across tested generations

DCas_v6 outperformed DCas_v5 by 77% in AMD STREAM memory bandwidth, 34% in Redis throughput, and 30% in SPEC CPU 2017. Against matched Das_v6 configurations, it retained 98% of AMD STREAM performance and 92% of Redis and SPEC CPU 2017 performance.

Microsoft's internal tests found DCas_v7 delivered 12% to 56% higher performance than DCas_v6 across 22 throughput and score results. Those preview tests used Linux and Windows guests at 8 and 32 virtual CPUs with SEV-SNP enabled on both generations. Fifth-generation AMD EPYC confidential VMs remain in gated preview.

Intel results vary by application and I/O path

Intel compared DCeds_v6 confidential VMs with Dds_v6 general-purpose VMs using matched 16-virtual-CPU, 64-GB configurations on fifth-generation Intel Xeon processors. Confidential VMs retained 97-99% of SPEC CPU performance, 97% for NGINX-WordPress, and 86% for PostgreSQL-HammerDB.

Local NVMe delivered nearly five times the input/output operations per second and approximately 16% lower average latency than the single-thread virtualized-SCSI baseline. At 32 threads, MANA networking reached general-purpose throughput parity and increased confidential VM throughput by 35% over the synthetic network path. Accelerated Networking through MANA is currently in preview for the Intel TDX configuration discussed.

SQL protection depends on encryption and recovery boundaries

SQL guidance covers encrypted query processing, driver behavior, and backup immutability. Each protects a different part of the data lifecycle, with specific limitations that matter when choosing a configuration.

Encrypted queries depend on the enclave threat model

Always Encrypted uses a 32-byte AES column encryption key protected by a column master key stored outside the server. Secure enclaves enable range comparisons and LIKE queries over randomized encrypted columns, while SQL Server 2022 added JOIN, GROUP BY, and ORDER BY operations inside secure enclaves.

Virtualization-based security enclaves are available across Azure SQL Database offerings in most regions, but currently lack enclave attestation. Intel SGX enclaves paired with Microsoft Azure Attestation provide the stronger protection matching the paper's threat model. Those options should not be treated as equivalent merely because both support enclave processing.

JDBC reuses eligible encryption-key cache entries

The Microsoft JDBC Driver for SQL Server now reuses eligible symmetric-key cache entries for enclave-enabled Always Encrypted key retrieval instead of repeating key-provider lookups. Other performance work reduces unnecessary ResultSet decoding and allocations while preserving lazy decoding and application-visible behavior.

An upcoming release will check cached authentication tokens before refresh synchronization and narrow synchronization scope to reduce contention. That authentication change is planned work, distinct from the encryption-key cache behavior described.

Backup immutability has a Hyperscale exclusion

The Microsoft SQL roundup describes automatic backup immutability protecting up to seven days of Azure SQL Database point-in-time restore backups. Hyperscale is currently excluded, so teams should verify coverage rather than assume the protection applies to every Azure SQL Database offering.

Other Security News

Azure Databricks credential-detection guidance correlates Entra ID risk detections with token logins within a one-hour window on either side, rather than relying on isolated workspace events. Microsoft Graph's risk detection API requires IdentityRiskEvent.Read.All and an Entra ID P1 or P2 license, while workspace Azure Monitor diagnostic settings omit account-level events and some services available through the audit system table.

Azure Local 2610 evaluates 429 security settings for domain-joined deployments and 403 for deployments using local identities, with CIS Level 1 alignment of 87% for both host types and 88% for the disconnected operations infrastructure VM. Excluding documented overrides raises alignment above 99%, but this is not certification or accreditation, and Set-AzsSecurity allows individual customization while retaining drift control across solution updates.

Gilles Chehade's backup discussion emphasizes reusing infrastructure inventories and repeatedly validating restores, because a completed backup job does not establish recoverability. His demonstration connected a reconciliation loop to GitHub CI to fail builds when expected backup coverage and actual state disagreed, using Plakar's backup capabilities, including end-to-end encryption and source-side deduplication.

Two .NET sessions provide separate development resources: Barry reviews security bugs from 2025-2026 and precautions against similar pitfalls, while the ASP.NET Core and Blazor session demonstrates .NET 11 application development. The latter covers asynchronous validation, richer API contracts, server-side rendering, Aspire integration, and building blocks for agentic user interfaces.

The Azure weekly video covers the ending of .NET 8 and .NET 9 support in Azure Static Web Apps and .NET 11 support in Azure Functions. Its networking coverage includes Azure Front Door web application firewall rate-limit grouping, alongside gRPC, edge actions, and Azure Route Server dual-stack support.

Thomas Maurer's sovereign AI session covers connected and fully disconnected Foundry Local deployments on Azure Local, along with region selection, the EU Data Boundary, confidential computing, and Purview controls. The article states that Microsoft announced general availability of Foundry Local on Azure Local for Microsoft Ignite 2026.

Also relevant this week

Dates to know

Deadlines