Shows how to build a daily dependency watcher using LangGraph and the Amazon Bedrock AgentCore Web Search API. The guide demonstrates enforcing date windows, source blocklists, and search budgets via the LangChain toolkit, ensuring these rules remain outside the model's control. It details using AWS SigV4 for authentication and AgentCore Gateway for centralized policy management.
Classified as: Feature
What to check
Verify the IAM permission bedrock-agentcore:InvokeWebSearch is granted to the execution role.
Configure the toolkit options exclude_domains, published_after, and max_searches to enforce fixed rules.
Check the account's Web Search quota (default 10 requests per second) to manage parallel branch throughput.
Tool / action control: The document describes using the LangChain toolkit to enforce date windows, blocklists, and search budgets in code, preventing the model from modifying them.
Human-in-the-loop support: The guide recommends human review of findings before action but does not describe a built-in approval workflow.
Identity and zero trust: The API uses AWS SigV4 signing with existing credentials and requires specific IAM permissions for the execution role.
Observability and audit: The document mentions LangGraph Studio for debugging and returning sources as data for checks, but does not detail comprehensive audit logging.
Data protection: The document states queries stay within the cloud environment and are served from an Amazon-operated index, but does not detail PII handling or residency.
Inter-agent communication security: AgentCore Gateway is mentioned as covering every agent and MCP client with one administrator source policy.
Donates OpenShell, an open source runtime for securing autonomous agents, to the CNCF Sandbox. OpenShell applies deterministic controls to agent activity, enforcing policies for tool use, inference access, and workload execution. It allows integration with existing compute, credential storage, and monitoring tools. NVIDIA also contributes to Agent Substrate for scaling agent workloads on Kubernetes.
Classified as: Feature
What to check
Verify that OpenShell deterministic controls are configured to restrict specific tool invocations and inference service access.
Confirm that the extensible architecture is integrated with existing credential storage and monitoring tools to enforce governance policies.
Tool / action control: The document states OpenShell applies deterministic controls to agent activity and enforces policies as agents use tools.
Identity and zero trust: The document mentions connecting credential storage and enforcing policies for access, but does not detail specific identity or zero-trust protocols.
Observability and audit: The document states teams can connect their own monitoring tools, but does not describe built-in tracing or audit features.
Adds InvokeWebSearch API to Amazon Bedrock AgentCore for general availability. The API returns ranked web search results with text extracts, titles, URLs, and publication times. It supports domain and time filters, IAM authorization, VPC access via PrivateLink, and CloudTrail logging. Billed at $7 per 1,000 queries. Available in six regions.
Classified as: Feature
What to check
Verify IAM policies restrict InvokeWebSearch API access to authorized principals.
Confirm CloudTrail data events are enabled to audit Web Search API queries.
Test domain and publication time filters to ensure search results align with trusted sources.
Tool / action control: The API allows filtering by domain and publication time, providing a mechanism to constrain search scope.
Identity and zero trust: The document explicitly states calls are authorized with AWS Identity and Access Management (IAM).
Observability and audit: The document states requests can be logged as AWS CloudTrail data events.
Data protection: The document mentions VPC access through AWS PrivateLink, which relates to network data protection, but does not detail PII handling or retention.
Adds bi-directional, zero-copy integration between AWS Analytics services and Workday Data Cloud. Teams can query Workday data via SageMaker Unified Studio, EMR, Glue, Redshift, or Athena without ETL pipelines. Access is governed by AWS and Workday permissions. Available in all commercial regions where AWS Glue is available.
Classified as: Feature
What to check
Verify that existing AWS and Workday permissions correctly restrict access to the new federated queries.
Confirm that the zero-copy integration does not bypass existing data residency or retention policies in either environment.
Not mentionedTool / action control
Not mentionedHuman-in-the-loop support
PartlyIdentity and zero trust
Not mentionedObservability and audit
PartlyData protection
N/AInter-agent communication security
Why these ratings
Identity and zero trust: Access is governed by existing AWS and Workday permissions, but no new zero-trust identity features are described.
Data protection: The integration retains existing security controls in both environments, implying data protection is maintained via current policies.
Introduces Control Room, a central system to track model workloads across research infrastructure. It enables rapid stopping of workloads when safety or security concerns arise. This capability is announced for OpenAI's internal research environment.
Classified as: Feature
What to check
Verify the scope of 'research infrastructure' to determine if Control Room applies to production or only internal R&D environments.
Confirm the specific triggers for 'safety or security concerns' that allow the system to stop workloads.
Adds evaluation form templates for customer service, inbound sales, and self-service scenarios. Templates assess human and AI agent behaviors and business outcomes like satisfaction and attrition risk. Managers can use or customize templates before deployment. Available in eight AWS Regions.
Classified as: Feature
What to check
Check the default questions in the customer service, inbound sales, and self-service templates.
Verify that managers can customize templates before deployment as stated.
Confirm the evaluation form templates are accessible in the eight listed AWS Regions.
Not mentionedTool / action control
PartlyHuman-in-the-loop support
Not mentionedIdentity and zero trust
PartlyObservability and audit
Not mentionedData protection
Not mentionedInter-agent communication security
Why these ratings
Human-in-the-loop support: Managers can customize templates before deploying, implying human oversight of the evaluation process.
Observability and audit: The feature enables automated evaluations of agent performance, contributing to observability.
Extends AgenticRetrieveStream with maxAgentIteration values 0 and 1. Value 0 returns passages without LLM calls; value 1 performs single-shot RAG. This allows applications to select retrieval effort per request via one API parameter.
Classified as: Feature
What to check
Verify that maxAgentIteration=0 correctly bypasses LLM calls to prevent unintended data processing.
Confirm that maxAgentIteration=1 executes exactly one retrieval and one generation call as documented.
Test that setting generateResponse=true with maxAgentIteration=0 returns a 4xx error as specified.
PartlyTool / action control
Not mentionedHuman-in-the-loop support
Not mentionedIdentity and zero trust
PartlyObservability and audit
Not mentionedData protection
Not mentionedInter-agent communication security
Why these ratings
Tool / action control: The maxAgentIteration parameter controls the scope of agent actions (planning rounds) and LLM invocation.
Observability and audit: Trace events are mentioned as showing planning and retrieval steps for inspection, but no specific audit logging controls are detailed.
Adds agentic mode to Azure Content Understanding for complex document tasks. This preview feature uses a reasoning loop with document-specific tools to navigate, inspect, and calculate results. It is positioned as a higher-cost, higher-accuracy alternative to standard mode for tasks requiring multi-step evidence synthesis.
Classified as: Feature
What to check
Verify the 'workflow': 'agentic' configuration in API version 2026-06-01-preview to ensure it is enabled only for intended complex tasks.
Test the additional compute cost and latency of agentic mode against standard mode using representative documents to justify usage.
Tool / action control: The document describes a 'document-native harness' with specific tools for navigating text, tables, and visual evidence.
Observability and audit: The document mentions 'validating results' and 'checking its work' as part of the reasoning loop, but does not detail external audit logging.
Adds hardware-isolated virtual machines for EKS pods, ECS tasks, and EC2 instances. Each workload runs in a Nitro-isolated VM with cryptographic attestation. Enables secure execution of untrusted or LLM-generated code on shared EC2 instances. Available on most nitro-based EC2 instances in all commercial regions at no additional charge.
Classified as: Feature
What to check
Verify that EKS pods or ECS tasks are configured to run on Nitro Sandboxes via the one-line change described.
Confirm that cryptographic attestation is enabled to ensure secrets are released only to proven-integrity sandboxes.
Test that LLM-generated code execution remains isolated within the Nitro VM boundary on shared EC2 instances.
Tool / action control: The document describes hardware isolation for executing untrusted code and LLM-generated code.
Observability and audit: States that CloudWatch works out of the box, implying standard logging availability, but does not describe specific agent tracing.
Data protection: Mentions support for EBS/EFS/S3 storage, implying data persistence capabilities, but does not specify PII handling or residency controls.
Graduates execution environments, A2A, and evaluation modules from experimental status. Adds multi-run evaluation and semantic tool argument matching. Introduces Microsoft Teams and OpenAI Decisions integrations. Changes Enum tool arguments to arrive as members and restricts RunSkillScriptTool to validated paths and arguments.
Classified as: SDK / product release
What to check
Verify Enum tool argument handling, as parameters now receive enum members instead of raw strings, breaking string-based lookups.
Update RunSkillScriptTool integrations to pass args, short_options, and positional_args, as free-form command strings are no longer accepted.
Review VertexAiSessionService.create_session calls, as extra keyword arguments are now limited to display_name, ttl, expire_time, and labels.
Tool / action control: RunSkillScriptTool now executes scripts from validated paths and arguments instead of free-form shell commands, and ContainerCodeExecutor supports input/output files.
Identity and zero trust: 26 auth modules graduate from experimental, but the release notes do not detail specific zero-trust identity changes for agents.
Observability and audit: Evaluation modules graduate from experimental, and telemetry updates include semconv bucket bounds for duration and token usage histograms.
Data protection: MongoDbMemoryService adds TTL expiration and vector search, but specific PII or residency controls are not detailed.
Inter-agent communication security: The A2A integration graduates from experimental to stable, and adk generate_agent_card builds an A2A agent card from the CLI.
Adds GPU partitioning to SageMaker Inference using NVIDIA Multi-Instance GPU (MIG). This allows dividing physical GPUs into isolated partitions with dedicated compute and memory. Teams can deploy more models per instance and select right-sized capacity fractions, removing the need for self-managed Kubernetes for MIG configuration.
Classified as: Feature
What to check
Verify that the selected AcceleratorPartitionLayouts are valid for the specific instance type (e.g., P4d, P5) to avoid endpoint creation rejection.
Check the NVIDIA driver version on P4d instances, as mig-1g.10gb requires R525 or newer, which may necessitate setting InferenceAmiVersion to al2023-ami-sagemaker-inference-gpu-4-1.
Identity and zero trust: The document mentions hardware-enforced tenant isolation and memory protection, which supports identity-based access boundaries, but does not detail agent identity or zero-trust authentication.
Data protection: Hardware-enforced memory isolation between partitions is cited as simplifying security and compliance reviews, implying data protection benefits.
Adds preview Data Cloud connectors enabling natural-language queries of BigQuery, Spanner, AlloyDB, Cloud SQL, and Looker. Executes queries under end-user identity with row- and column-level security. Supports federated query mode without data copying. Integrates with Knowledge Catalog for semantic discovery and context mapping.
Classified as: Feature
What to check
Verify that the default read-only connection setting in the Gemini admin console aligns with your data access policies.
Confirm that Knowledge Catalog metadata visibility is restricted to users with appropriate permissions to prevent data leakage via semantic discovery.
Adds agent trace support to the Amazon OpenSearch Service optimized engine. Maps OpenTelemetry logs and traces to object, nested, and flat object fields. Supports update, delete, and upsert operations on hot-tier data. Available on OpenSearch 3.7 in twelve AWS Regions at no additional charge.
Classified as: Feature
What to check
Verify that OpenTelemetry data sent to the domain is correctly mapped to object, nested, and flat object fields.
Confirm that update, delete, and upsert operations on hot-tier data function as expected for late-arriving or erroneous data.
Test the Discover Traces, service map, and agent traces views in the OpenSearch UI to ensure accurate visualization of agent sessions.
Changes the default Gemini Developer API version to v1 and the default HTTP client to httpx2. Updates CodeExecutionResultStep.result to a union type. Adds source_text to AudioTranscriptionConfig. These are breaking changes requiring code or environment variable updates.
Classified as: SDK / product release
What to checkVerify if relying on v1beta defaults for Gemini Developer API; set GOOGLE_GENAI_API_VERSION=beta to restore. Install google-genai[httpx] and set GOOGLE_GENAI_HTTP_CLIENT=httpx to revert HTTP client. Update code to handle CodeExecutionResultStep.result as a union type.
Not mentionedTool / action control
Not mentionedHuman-in-the-loop support
Not mentionedIdentity and zero trust
Not mentionedObservability and audit
PartlyData protection
Not mentionedInter-agent communication security
Why these ratings
Data protection: The switch to httpx2 changes the default SSL certificate verification to use the OS trust store, affecting data transmission security configuration.
Adds Web Answer, a managed tool returning synthesized, cited answers to natural-language questions. It extends Web Search by planning parallel searches and synthesizing evidence. Agents attach it via AgentCore Gateway using MCP and IAM authentication. Domain include/deny lists and publish-date filters scope search boundaries. Available in US East (N. Virginia) with per-tool call pricing.
Classified as: Feature
What to checkConfigure domain include/deny lists and publish-date filters to define the search scope for Web Answer.
Adds support for IAM Identity Center multi-region replication in Amazon SageMaker AI Studio. This allows deploying SageMaker AI Studio domains in regions different from the primary IAM Identity Center instance. The change supports data residency and compliance requirements while maintaining centralized workforce identity management and single sign-on capabilities.
Classified as: Feature
What to check
Verify that the primary IAM Identity Center instance is replicated to the target region before deploying SageMaker AI Studio domains.
Confirm that the new multi-region configuration aligns with organizational data residency and compliance mandates.
Identity and zero trust: The item directly addresses identity management by enabling multi-region replication of IAM Identity Center for centralized workforce identity.
Data protection: The document states the capability helps satisfy data sovereignty and residency requirements, but does not detail specific data protection controls.
Adds local inference for Qwen3.8-27B in Foundry Local, connecting to GitHub Copilot. Requires Foundry Local CLI 0.13.0 or later and NVIDIA GPUs with at least 24 GB VRAM. The document states code and context remain on the local machine.
Classified as: Feature
What to check
Check that the 'Base URL' field in GitHub Copilot settings is configured with the specific API endpoint returned by 'foundry server status'.
Verify that the installed Foundry Local CLI version is 0.13.0 or higher to ensure the described optimizations are active.
Not mentionedTool / action control
Not mentionedHuman-in-the-loop support
Not mentionedIdentity and zero trust
Not mentionedObservability and audit
PartlyData protection
Not mentionedInter-agent communication security
Why these ratings
Data protection: The source states that local inference keeps code and context on the machine, but this is a deployment characteristic rather than a described governance control.
Adds Australia, India, and Japan residency options; adds GPT-6 Rosalind to the model catalog; adds image complexity and GIF/SVG output controls; expands agent turn errors and image/audio options.
Classified as: SDK / product release
What to check
Check API calls for new Australia, India, and Japan residency options to verify data location compliance.
Verify model catalog entries for GPT-6 Rosalind to confirm availability and usage policies.
Not mentionedTool / action control
Not mentionedHuman-in-the-loop support
Not mentionedIdentity and zero trust
PartlyObservability and audit
DirectlyData protection
Not mentionedInter-agent communication security
Why these ratings
Observability and audit: The release notes state it expands agent turn errors, which may affect error logging and tracing.
Data protection: The release notes explicitly add Australia, India, and Japan residency options.
Describes an autonomous pipeline that identified, validated, and generated patches for 948 local privilege-escalation vulnerabilities in the Linux kernel. This research demonstrates automated security remediation capabilities for critical system software.
Classified as: Research
What to check
Confirm the pipeline's capability to generate patches for local privilege-escalation vulnerabilities as stated in the research.
Verify the count of 948 vulnerabilities identified and patched in the Linux kernel.
Adds error detail types and an IDs filter for listing workspaces. Introduces specific error classes for HTTP statuses and API codes. Fixes Bedrock path validation to reject invalid model names. Ensures the memory tool preserves the /memories directory. Updates documentation for plugin reach and telemetry headers.
Classified as: SDK / product release
What to check
Verify that Bedrock integrations handle the new ValueError exceptions for empty or invalid model names (e.g., '.', '..') to prevent unintended path requests.
Confirm that applications relying on the memory tool do not depend on the previous behavior of renaming or deleting the /memories directory, as this is now prevented.
Tool / action control: The release fixes a bug where the memory tool could rename or delete the /memories directory, addressing a specific tool behavior risk.
Observability and audit: The release adds error detail types and clarifies telemetry destination headers in documentation, enhancing error visibility.
Adds Smart Approvals to Amazon Quick, which uses a classifier to automatically approve eligible actions or return them for user review. It operates within existing Custom Permissions and Guardrails, is on by default, and records permission updates in AWS CloudTrail.
Classified as: Feature
What to checkVerify that connector action permissions are configured to mandate human approval for high-impact actions like external sends or deletions. Confirm that Custom Permissions are set to allow or deny Smart Approvals for specific user profiles, noting it is on by default. Ensure a CloudTrail trail is configured to retain permission update events beyond the default 90-day history.
Tool / action control: The feature evaluates actions against Custom Permissions, Quick Guardrails, and connector action permissions before automated review.
Human-in-the-loop support: It is designed to preserve human oversight by returning high-risk or ineligible actions to the user for review.
Identity and zero trust: Administrators use Custom Permissions to determine who can use the feature, and approvals are single-use and specific to the action.
Observability and audit: Permission updates and verdicts are recorded in AWS CloudTrail, and chat logs are delivered to Amazon CloudWatch Logs.
Data protection: The document states that Quick does not use classifier review content to train models and advises masking sensitive data in logs.
Adds a guidelines array to Chat Completions or Responses API requests to set content handling for nsfw, violence, and profanity. Grok follows these guidelines during response generation. This feature is documented in the release notes dated October 27.
Classified as: Feature
What to check
Check the guidelines array parameter in Chat Completions or Responses API requests.
Verify the nsfw, violence, and profanity category settings are applied as intended.
Not mentionedTool / action control
Not mentionedHuman-in-the-loop support
Not mentionedIdentity and zero trust
Not mentionedObservability and audit
PartlyData protection
Not mentionedInter-agent communication security
Why these ratings
Data protection: The feature allows configuring content guidelines for sensitive categories like nsfw and violence, which relates to data protection policies.
Adds bedrock-runtime endpoint support for OpenAI Daybreak Red: GPT-5.6 Cyber and Daybreak Blue: GPT-5.6 Sol. These models are available via US and Global cross-Region inference profiles. Access requires enrollment in Daybreak access from OpenAI. Inference data is not used for model training, and data sharing with OpenAI is not required.
Classified as: Model release
Governance ratings: N/A (not applicable to a model release)
Releases MiniMax-H3-Image-PE, a prompt expansion model fine-tuned from Qwen3.8-27B. It converts short text requests into detailed image descriptions and aspect-ratio recommendations. Available via Hugging Face and MiniMax API for text-to-image and image editing workflows.
Adds NVIDIA Nemotron 3.5 Super VL to Amazon SageMaker JumpStart. The 120B-parameter vision-language model supports image, video, and text inputs with up to 1M token context. Available in NVFP4 and BF16 variants for deployment on G7 or P5en instances.
Classified as: Model release
Governance ratings: N/A (not applicable to a model release)
Adds Claude Fable 5.2 and Claude Mythos 5.2 to Amazon Bedrock. Fable 5.2 is generally available for technical work; Mythos 5.2 is restricted to verified cyber security and life sciences use. Fable 5.2 requires retention for up to 30 days and human review by default, or zero data retention until June 30, 2027.
Adds Kimi K3.1 Pro (Highspeed) to Amazon Bedrock. This open-weight, natively multimodal agentic model features a 1-million-token context window and 3-trillion-parameter-class architecture. It is available in all AWS Regions where Amazon Bedrock is available via cross-Region inference.
Releases Gemini Omni 1.2 Flash for video generation and editing. Supports 15-second clips, with drafts at 360p or 720p upscaled to 1080p or 4K. Priced at $0.07 per second for 720p video. Available via API and Google Vids; coming soon to Gemini for business.
Classified as: Model release
Governance ratings: N/A (not applicable to a model release)
Introduces Images 2.6, a model update offering faster image generation, GIF creation, and precise editing. Available across ChatGPT, Codex, and the API with new controls.
Classified as: Model release
Governance ratings: N/A (not applicable to a model release)
Releases Kimi-K3.1-Pro, an open-weight, natively multimodal agentic model with 2.8T total parameters and a 1-million-token context window. The model is available on Hugging Face under the moonshotai organization. It supports text and image modalities and is designed for agentic capabilities including coding and research.
AMD 6th Gen EPYC SP7 processors are generally available. Systems ship from leading server vendors. Cloud instances on Microsoft Azure and OCI are generally available soon. The portfolio includes AMD Helion agentic AI racks with up to 73,728 threads per rack.
Classified as: AI hardware
Governance ratings: N/A (not applicable to AI hardware)
AMD EPYC 5th Gen processors serve as the host CPU for Google Cloud TPU 8tx, enabling hardware-enforced confidential computing. This configuration protects sensitive AI models and data during inference, fine-tuning, and reinforcement learning workloads by extending AMD Secure Encrypted Virtualization (SEV) capabilities to connected accelerators.
Classified as: AI hardware
Governance ratings: N/A (not applicable to AI hardware)