ANNOUNCED
NVIDIA Release
· 2026-10-15
NVIDIA announced Nemotron 3.5 Super VL, a 120B-parameter MoE model activating 12B parameters per token. It combines reasoning, tool use, and multimodal understanding. Teams can post-train it using NVIDIA NeMo for specialized agents, claiming up to 5x higher throughput and 40x lower inference cost in specific use cases.
What to checkVerify tool permissions and sandboxing when deploying this model for agentic tasks. Ensure local deployment configurations meet data residency requirements. Monitor tool call logs for unexpected behavior.
PartlyTool / action control
Not addressedHuman-in-the-loop support
Not addressedIdentity and zero trust
PartlyObservability and audit
PartlyData protection
Not addressedInter-agent communication security
Why these scores
- Tool / action control: The model supports tool use and post-training for tool interfaces, but specific allowlisting or sandboxing controls are not described.
- Human-in-the-loop support: The document describes autonomous agent workflows and does not mention human-in-the-loop approval mechanisms.
- Identity and zero trust: No details on agent identity management, authentication, or least privilege enforcement are provided.
- Observability and audit: The model is evaluated on benchmarks and supports tracing via agent harnesses, but specific audit logging features are not detailed.
- Data protection: The model supports local deployment and open weights for data residency, but specific PII handling or retention policies are not defined.
- Inter-agent communication security: The document focuses on single-model capabilities and does not address security protocols for inter-agent communication.
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16
ANNOUNCED
Meta Release
· 2026-10-12
Meta introduced Muse Wonder and Muse Wonder Full, models for agentic tasks like insurance claims and coding. They feature Active Memory for cross-conversation context and Active Focus for error recovery. Availability is via Meta Model API and Muse Code, with a future rollout to Muse subscribers.
What to checkVerify the 'Active Memory' feature against your data retention policies to ensure PII is not persisted indefinitely. Implement human-in-the-loop gates for high-stakes actions, as the model's 'checking in' behavior is a model-level feature, not a system-level control.
Not addressedTool / action control
PartlyHuman-in-the-loop support
Not addressedIdentity and zero trust
Not addressedObservability and audit
PartlyData protection
Not addressedInter-agent communication security
Why these scores
- Tool / action control: The document does not describe specific tool allowlisting, sandboxing, or permission controls.
- Human-in-the-loop support: The model is described as checking in before following through on important actions, but no specific HITL workflow is detailed.
- Identity and zero trust: No details are provided regarding agent identity management, authentication, or least privilege enforcement.
- Observability and audit: The text mentions safety evaluations but does not describe tracing, logging, or audit capabilities for agents.
- Data protection: The model is trained to avoid sharing private information without permission, but specific data residency or retention policies are not stated.
- Inter-agent communication security: The document focuses on single-agent capabilities and does not address security for communication between multiple agents.
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16
ANNOUNCED
AWS Release
· 2026-10-13
AWS announced Grok 4.7 on Amazon Bedrock, offering configurable reasoning effort levels (low, medium, high, xhigh) and cross-region inference profiles. The model supports tool calling and structured output, enabling teams to tune agent deliberation and verify performance on complex, multi-step tasks.
What to checkConfigure IAM policies for specific inference profiles and use short-term bearer tokens. Monitor token usage and reasoning effort levels to balance cost and accuracy.
DirectlyTool / action control
Not addressedHuman-in-the-loop support
DirectlyIdentity and zero trust
DirectlyObservability and audit
PartlyData protection
Not addressedInter-agent communication security
Why these scores
- Tool / action control: The source describes constraining tool parameters with enums and using a standard tool-calling loop to execute returned calls.
- Human-in-the-loop support: The document does not describe any human-in-the-loop approval or intervention mechanisms for the agent.
- Identity and zero trust: The source recommends generating short-term bearer tokens from existing AWS credentials to tie authentication to IAM identity.
- Observability and audit: The document details tracking prompt and completion tokens across the agent loop and monitoring reasoning effort levels.
- Data protection: The source mentions US data residency requirements for the us.xai.grok-4.7 inference profile but does not detail PII handling.
- Inter-agent communication security: The document focuses on a single agent interacting with tools, not on communication between multiple agents.
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16
ANNOUNCED
AWS Feature
· 2026-10-13
AWS describes two methods to automate AWS Marketplace Private Pricing Offer activation for Amazon Bedrock. A lightweight script handles one-time trusted license activation, while a CDK-based solution uses EventBridge and Lambda for ongoing, allowlisted grant distribution across AWS Organizations.
What to checkTreat the seller allowlist as a strict security boundary; verify every seller account ID against accepted private offers before seeding. Use the dry-run mode of the lightweight script as a mandatory review gate before applying changes to production accounts.
PartlyTool / action control
PartlyHuman-in-the-loop support
DirectlyIdentity and zero trust
DirectlyObservability and audit
Not addressedData protection
Not addressedInter-agent communication security
Why these scores
- Tool / action control: The solution uses a seller allowlist to restrict which third-party models can be activated, but this is a procurement control rather than runtime action sandboxing.
- Human-in-the-loop support: The lightweight path requires manual review and confirmation, while the automated path relies on pre-configured allowlists rather than real-time human approval.
- Identity and zero trust: The item directly addresses identity and privilege by managing distributed grants and scoping access to specific AWS Organizations, OUs, or accounts.
- Observability and audit: The CDK path explicitly logs every activation attempt or skip decision to Amazon CloudWatch Logs for auditability.
- Data protection: The document focuses on pricing and access management, with no mention of PII handling, data residency, or retention policies.
- Inter-agent communication security: The item concerns infrastructure provisioning and billing, not the security of communication between AI agents.
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16
ANNOUNCED
AWS Feature
· 2026-10-13
Anchor Browser, a third-party solution on AWS Marketplace, integrates with Amazon Bedrock AgentCore payments. This allows agents to handle HTTP 402 challenges by moving wallet keys and budget controls to managed infrastructure, with observability via CloudWatch.
What to check- Configure strict spending limits in AgentCore payments.
- Ensure wallet credentials are stored in AgentCore Identity, not the agent runtime.
- Monitor payment events in CloudWatch for audit compliance.
DirectlyTool / action control
PartlyHuman-in-the-loop support
DirectlyIdentity and zero trust
DirectlyObservability and audit
PartlyData protection
Not addressedInter-agent communication security
Why these scores
- Tool / action control: The source describes enforcing configurable spending limits at the infrastructure layer and moving sensitive operations out of the agent runtime.
- Human-in-the-loop support: The document mentions observability into 'approvals' and 'denials' but describes the workflow as autonomous without detailing specific human intervention gates.
- Identity and zero trust: The source details the use of IAM roles, workload identities in AgentCore Identity, and secure credential storage to manage agent identity and access.
- Observability and audit: The document states that AgentCore payments provides observability into payment attempts, approvals, denials, and settlements through Amazon CloudWatch.
- Data protection: The source mentions secure storage of wallet credentials and keys but does not detail PII handling, residency, or retention policies.
- Inter-agent communication security: The document describes agent-to-merchant payment flows but does not address security protocols for communication between multiple agents.
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16
ANNOUNCED
AWS Feature
· 2026-10-14
AWS describes a solution using Amazon Bedrock AgentCore runtime and Strands Agents framework to generate PowerPoint presentations. It utilizes Anthropic Claude Sonnet 4.6, Amazon Titan Text Embeddings v2, and Stability AI Stable Image Inpaint. The architecture includes RAG via Amazon OpenSearch Serverless and security controls like IAM and Guardrails.
What to checkVerify that the 'built-in identity' and IAM roles enforce least privilege for the agent's tool invocations. Ensure Guardrails policies are tested against specific PII and injection scenarios before production use.
DirectlyTool / action control
Not addressedHuman-in-the-loop support
DirectlyIdentity and zero trust
DirectlyObservability and audit
DirectlyData protection
Not addressedInter-agent communication security
Why these scores
- Tool / action control: The document describes a supervisor agent routing to specific tools (generate_pptx, generate_summary) and mentions sandboxing via isolated containers.
- Human-in-the-loop support: The document describes an automated workflow completing in under 90 seconds without mentioning human approval steps.
- Identity and zero trust: The solution explicitly uses IAM least-privilege access control, AWS STS role assumption per tenant, and built-in identity for the agent.
- Observability and audit: The document states the agent code runs with built-in observability and mentions AWS X-Ray tracing on Lambda functions.
- Data protection: The document details PII handling via Guardrails, KMS encryption, and S3 versioning/logging for data protection.
- Inter-agent communication security: The architecture describes a single supervisor agent interacting with tools, not multiple agents communicating with each other.
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16
ANNOUNCED
AWS Feature
· 2026-10-14
AWS describes a multi-agent system on Bedrock AgentCore that classifies and extracts data from documents. It uses Strands Agents, Textract, and S3 Vectors. The system learns extraction rules from new formats and reuses them for known types, reducing manual template configuration.
What to checkVerify that AgentCore Identity enforces least-privilege access for each agent role. Implement logging of all MCP tool calls to audit agent actions. Review the 'Instructions' stored in the vector store to ensure no sensitive data is persisted in learned prompts.
PartlyTool / action control
PartlyHuman-in-the-loop support
DirectlyIdentity and zero trust
DirectlyObservability and audit
Not addressedData protection
Not addressedInter-agent communication security
Why these scores
- Tool / action control: Agents access AWS resources through machine-to-machine credentials and MCP interfaces, implying scoped permissions, but specific allowlisting details are not described.
- Human-in-the-loop support: The system includes a chat interface for troubleshooting and a review tab for verifying extracted data, but it is designed for minimal human intervention.
- Identity and zero trust: The document explicitly states agents securely access AWS resources through machine-to-machine credentials managed by AgentCore Identity and Amazon Cognito.
- Observability and audit: The solution uses AgentCore Observability and Amazon CloudWatch telemetry to gain visibility into agent performance and status.
- Data protection: The document does not address PII handling, data residency, or retention policies for the processed documents.
- Inter-agent communication security: The document describes the workflow graph but does not specify security controls for communication between the eight specialized agents.
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16
ANNOUNCED
AWS Feature
· 2026-10-15
AWS introduces Trust in Depth, a layered governance approach for agents. It combines automated reasoning, identity management via AgentCore, and observability through CloudWatch Omni to enforce boundaries and audit actions, aiming to reduce reliance on manual human approval while maintaining security.
What to checkDefine precise agent scopes and policies before deployment. Implement automated reasoning checks for rule compliance. Use centralized registries to track agent identities and ensure all actions are traceable.
DirectlyTool / action control
DirectlyHuman-in-the-loop support
DirectlyIdentity and zero trust
DirectlyObservability and audit
PartlyData protection
PartlyInter-agent communication security
Why these scores
- Tool / action control: The document describes Bedrock AgentCore Policies and Strands Box for enforcing tool and action boundaries.
- Human-in-the-loop support: The framework explicitly reserves human judgment for ambiguous decisions and high-stakes actions.
- Identity and zero trust: AgentCore issues unique identities to agents and checks them on every request to ensure accountability.
- Observability and audit: CloudWatch Omni is described as capturing traces and evaluating agent behavior for correctness and faithfulness.
- Data protection: Guardrails are mentioned for redacting sensitive data, but specific residency or retention controls are not detailed.
- Inter-agent communication security: The Agent Registry manages MCP servers and tools, implying secure discovery, but direct inter-agent communication protocols are not specified.
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16
ANNOUNCED
AWS Release
· 2026-10-16
Reflection AI introduced Beam, a 501 billion parameter open-weight model for agentic coding. It is available via SageMaker JumpStart in US East regions. Teams can deploy it using managed endpoints or Kubernetes-native HyperPod Inference with SGLang or vLLM runtimes.
What to check- Verify IAM permissions for the new SageMaker endpoints.
- Review the Apache 2.0 license terms for open-weight usage.
- Assess GPU capacity requirements for the 501B parameter model.
Not addressedTool / action control
Not addressedHuman-in-the-loop support
PartlyIdentity and zero trust
PartlyObservability and audit
Not addressedData protection
Not addressedInter-agent communication security
Why these scores
- Tool / action control: The document describes model deployment infrastructure but does not specify tool allowlisting or action control mechanisms.
- Human-in-the-loop support: No human-in-the-loop approval workflows or intervention points are described in the deployment guide.
- Identity and zero trust: The document mentions IAM permissions and isolated Hugging Face tokens, but does not detail agent-specific identity or zero-trust architecture.
- Observability and audit: The text references observability in the context of HyperPod control but does not detail specific tracing, logging, or audit capabilities for agent actions.
- Data protection: The document focuses on model weights and deployment configuration without addressing PII handling, residency, or retention policies.
- Inter-agent communication security: The source describes single-model deployment and does not address security protocols for communication between multiple agents.
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16
ANNOUNCED
AWS Research
· 2026-10-16
AWS announced a nine-stage evaluation framework using Amazon Bedrock AgentCore to monitor agent quality, safety, and performance. This addresses silent degradation in multi-agent systems by integrating evaluation into CI/CD and production monitoring.
What to check- Implement continuous evaluation gates in CI/CD pipelines.
- Define quality metrics and golden sets for each agent.
- Monitor for silent degradation in production.
Governance areas it touches: Tool / action control · Observability and audit · Inter-agent communication security
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16
ANNOUNCED
AWS Feature
· 2026-10-15
AWS announced that the next generation of AWS Resilience Hub integrates with AWS DevOps Agent. This allows teams to review failure mode assessment findings and resilience test results. The agent performs structured analysis of test observations and supports on-demand investigation tasks to create mitigation plans.
What to checkConfigure the DevOps Agent within the Resilience Hub service settings to define its scope. Review the structured analysis outputs to validate root cause findings before implementing mitigations.
PartlyTool / action control
DirectlyHuman-in-the-loop support
Not addressedIdentity and zero trust
DirectlyObservability and audit
Not addressedData protection
Not addressedInter-agent communication security
Why these scores
- Tool / action control: The agent is configured in the service configuration, implying permission scoping, but specific allowlisting details are not described.
- Human-in-the-loop support: The source explicitly states users can start the agent on demand and ask questions to investigate findings.
- Identity and zero trust: The document does not mention agent identity, authentication, or least privilege mechanisms.
- Observability and audit: The agent returns structured analysis of test observations, including symptoms, root cause, and impact.
- Data protection: No specific details on PII handling, data residency, or retention are provided in the text.
- Inter-agent communication security: The document describes a single agent interacting with a service, not communication between multiple agents.
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16
ANNOUNCED
AWS Feature
· 2026-10-15
Amazon MWAA integrates with the Agent Toolkit for AWS to provide four built-in AI agent skills. These skills allow AI coding assistants to author, debug, test, and upgrade MWAA workflows using natural language. The feature supports both Provisioned and Serverless environments in all AWS Regions where MWAA is available.
What to check- Verify the identity and permissions of the AI coding assistant before enabling skills.
- Review the generated DAGs and code changes manually before deployment.
- Ensure the aws-data-analytics plugin is installed from a trusted source.
PartlyTool / action control
Not addressedHuman-in-the-loop support
Not addressedIdentity and zero trust
Not addressedObservability and audit
Not addressedData protection
Not addressedInter-agent communication security
Why these scores
- Tool / action control: The item introduces specific agent skills that require configuration and plugin installation to control tool usage.
- Human-in-the-loop support: The document describes autonomous interaction with MWAA but does not mention human approval gates or review steps.
- Identity and zero trust: The source does not specify how agent identity, authentication, or least privilege is handled for these skills.
- Observability and audit: No details are provided regarding logging, tracing, or audit capabilities for agent actions.
- Data protection: The document does not address PII handling, data residency, or retention policies for agent interactions.
- Inter-agent communication security: The item focuses on single-agent interaction with MWAA and does not describe security for inter-agent communication.
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16
ANNOUNCED
AWS Feature
· 2026-10-15
AWS announced the availability of Meta's Muse Spark Open model on SageMaker JumpStart. This 816B parameter MoE model supports 1M-token contexts and agentic coding workflows. It is offered in BF16 and NVFP4 variants for deployment within customer AWS accounts.
What to check- Verify IAM permissions for the SageMaker endpoint.
- Review data residency settings for the 1M-token context window.
- Assess tool-calling permissions before enabling agentic workflows.
Not addressedTool / action control
Not addressedHuman-in-the-loop support
PartlyIdentity and zero trust
Not addressedObservability and audit
PartlyData protection
Not addressedInter-agent communication security
Why these scores
- Tool / action control: The document describes model capabilities but does not specify tool allowlisting or sandboxing controls.
- Human-in-the-loop support: No human-in-the-loop approval mechanisms or intervention points are described in the source.
- Identity and zero trust: Deployment in the customer's own AWS account implies standard IAM identity controls apply, but specific agent identity features are not detailed.
- Observability and audit: The text mentions deployment methods but does not describe specific tracing, logging, or audit features for agent actions.
- Data protection: Running in the customer's account suggests data residency control, but specific PII handling or retention policies are not stated.
- Inter-agent communication security: The document focuses on single-model deployment and does not address security for communication between multiple agents.
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16
ANNOUNCED
AWS Release
· 2026-10-15
AWS announces general availability of the Go SDK for Lambda durable functions. This enables Go developers to build long-running workflows with automatic checkpointing, human/agent-in-the-loop callbacks, and local testing via durabletest and durablelint tools.
What to checkIntegrate durable function checkpoints into existing audit logs. Define clear approval gates for human-in-the-loop callbacks to prevent unauthorized workflow progression.
Not addressedTool / action control
DirectlyHuman-in-the-loop support
Not addressedIdentity and zero trust
PartlyObservability and audit
Not addressedData protection
Not addressedInter-agent communication security
Why these scores
- Tool / action control: The document describes workflow execution mechanics but does not mention tool allowlisting or sandboxing controls.
- Human-in-the-loop support: The source explicitly states the SDK includes 'callback integration for human and agent-in-the-loop workflows' and supports pausing for human review.
- Identity and zero trust: No details regarding agent identity management, authentication, or least privilege enforcement are provided in the text.
- Observability and audit: The SDK provides 'steps for progress tracking' and local testing tools, but does not describe centralized audit logging or tracing capabilities.
- Data protection: The document does not address PII handling, data residency, or retention policies for the durable functions.
- Inter-agent communication security: While it mentions 'agent-in-the-loop', it does not describe security protocols for inter-agent communication.
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16
ANNOUNCED
AWS Feature
· 2026-10-15
AWS announced AWS Transform for migrations is available in Asia Pacific (Singapore). This provides data residency within the region and lower latency for local infrastructure. The service uses agentic AI to automate migration lifecycles for workloads from VMware, Hyper-V, KVM, bare metal, and hybrid environments.
What to check- Verify data residency compliance for Singapore workloads.
- Review agent permissions and audit logs before enabling automation.
Not addressedTool / action control
PartlyHuman-in-the-loop support
Not addressedIdentity and zero trust
Not addressedObservability and audit
DirectlyData protection
Not addressedInter-agent communication security
Why these scores
- Tool / action control: The document does not describe specific tool allowlisting or sandboxing controls.
- Human-in-the-loop support: The service is described as a 'collaborative' workbench, implying human oversight, but specific HITL mechanisms are not detailed.
- Identity and zero trust: No specific agent identity or zero-trust authentication details are provided.
- Observability and audit: The text does not mention tracing, logging, or audit capabilities for the agents.
- Data protection: The source explicitly states the service gives customers 'data residency within the region'.
- Inter-agent communication security: No details on secure communication protocols between agents are provided.
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16
ANNOUNCED
AWS Feature
· 2026-10-15
AWS announced Web Search on Amazon Bedrock AgentCore is available in US East (Ohio), US West (Oregon), Asia Pacific (Sydney), and AWS GovCloud (US-West). This expansion supports lower latency and data residency for government workloads.
What to check- Verify data residency policies align with the new region availability.
- Confirm compliance requirements for government workloads are met by GovCloud deployment.
Not addressedTool / action control
Not addressedHuman-in-the-loop support
Not addressedIdentity and zero trust
Not addressedObservability and audit
PartlyData protection
Not addressedInter-agent communication security
Why these scores
- Tool / action control: The document describes regional availability but does not mention allowlisting, permissions, or sandboxing controls.
- Human-in-the-loop support: No human-in-the-loop mechanisms or approval workflows are described in the source.
- Identity and zero trust: The text does not address agent identity, authentication, or least privilege configurations.
- Observability and audit: Tracing, logging, or evaluation capabilities are not mentioned in the announcement.
- Data protection: The source states GovCloud availability helps meet compliance and data residency requirements.
- Inter-agent communication security: The document does not discuss security protocols for communication between agents.
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16
SHIPPED
Google Release
· 2026-10-13
Google ADK v2.12.0 introduces after_tool_callback to filter tool outputs and require_confirmation for workflow nodes. It also supports dedicated service accounts for Agent Engine deployment and Python 3.15 compatibility.
What to checkConfigure after_tool_callback to sanitize sensitive data in tool responses. Enable require_confirmation for high-risk workflow steps. Assign unique service accounts to each agent deployment to enforce least privilege.
DirectlyTool / action control
DirectlyHuman-in-the-loop support
DirectlyIdentity and zero trust
PartlyObservability and audit
PartlyData protection
PartlyInter-agent communication security
Why these scores
- Tool / action control: The release adds after_tool_callback to intercept and filter tool outputs before returning them to the model.
- Human-in-the-loop support: The release introduces require_confirmation to mandate human approval before executing workflow nodes.
- Identity and zero trust: The release enables deploying agents to Agent Engine using dedicated custom service accounts.
- Observability and audit: The release preserves model thinking parts and sub-agent hierarchy in ADK events, aiding traceability.
- Data protection: The release mentions continuing with de-identified text, implying data handling controls.
- Inter-agent communication security: The release adds skill-level security support to AgentCardBuilder, affecting agent interaction metadata.
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16
SHIPPED
Google Release
· 2026-10-14
Google Gen AI SDK v2.31.0 adds gemini-4-argon to the Model enum. It fixes function declaration parsing for union types, corrects grounding metadata merging, and improves continuation stream retry logic. It also aligns mTLS endpoints with transport SSL contexts.
What to checkVerify mTLS configuration for all transports. Test function declaration parsing with union types. Monitor continuation stream retries for stability.
PartlyTool / action control
Not addressedHuman-in-the-loop support
PartlyIdentity and zero trust
Not addressedObservability and audit
Not addressedData protection
Not addressedInter-agent communication security
Why these scores
- Tool / action control: Fixes function declaration parser for union types, improving tool schema handling.
- Human-in-the-loop support: No specific human-in-the-loop controls are mentioned in the changelog.
- Identity and zero trust: Matches mTLS endpoint to each transport's SSL context, enhancing transport security.
- Observability and audit: No new tracing, logging, or audit features are described.
- Data protection: No specific data protection, PII, or residency changes are noted.
- Inter-agent communication security: No inter-agent communication security features are addressed.
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16
ANNOUNCED
Google Release
· 2026-10-13
Google announced Gemini 4 Argon, a frontier model for multi-step agentic tasks. It features improved resilience against indirect prompt injection and cyber mitigations. Access is available via API and Gemini for business, with introductory pricing through November 30, 2026.
What to checkVerify the model's prompt injection resilience in your specific threat model. Ensure your existing tool allowlisting aligns with Argon's cyber mitigations. Monitor the introductory pricing window for budget planning.
PartlyTool / action control
PartlyHuman-in-the-loop support
Not addressedIdentity and zero trust
Not addressedObservability and audit
Not addressedData protection
Not addressedInter-agent communication security
Why these scores
- Tool / action control: The model includes cyber mitigations blocking offensive tasks and pauses for confirmation before irreversible steps.
- Human-in-the-loop support: Argon is trained to pause for confirmation before taking irreversible steps.
- Identity and zero trust: The document does not describe specific agent identity, authentication, or least-privilege mechanisms.
- Observability and audit: No specific tracing, logging, or audit capabilities are detailed in the announcement.
- Data protection: The text mentions regulated work but does not specify PII handling, residency, or retention controls.
- Inter-agent communication security: The document focuses on single-model resilience, not secure communication between distinct agents.
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16
ANNOUNCED
Google Feature
· 2026-10-14
Google announced general availability of token-based cost management for BigQuery generative AI functions. Administrators can set daily input and output token quotas at project and user levels via IAM. This prevents budget overruns by stopping LLM calls when limits are reached, distinct from dollar-based spend caps.
What to checkConfigure per-user token quotas to isolate experimental workloads from production budgets. Use AI.COUNT_TOKENS to estimate costs before scaling large datasets. Monitor usage trends to adjust limits as model pricing or workload volume changes.
PartlyTool / action control
Not addressedHuman-in-the-loop support
PartlyIdentity and zero trust
DirectlyObservability and audit
Not addressedData protection
Not addressedInter-agent communication security
Why these scores
- Tool / action control: The item provides usage limits (quotas) for AI functions but does not describe allowlisting or sandboxing of specific tools.
- Human-in-the-loop support: The document focuses on automated quota enforcement and does not describe human approval workflows for agent actions.
- Identity and zero trust: Quotas are managed via IAM and support per-user limits, but the item does not address agent identity or zero-trust authentication mechanisms.
- Observability and audit: The source describes monitoring historical token consumption and detecting spikes via the Quotas & System Limits page.
- Data protection: The item addresses cost management and does not mention PII handling, data residency, or retention policies.
- Inter-agent communication security: The item concerns single-user or project-level BigQuery queries and does not address communication between agents.
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16
ANNOUNCED
Google Research
· 2026-10-16
Google Cloud published four lessons for evaluating agent plugins: grade execution trajectories, preserve turn structure in telemetry, run live and plan-only tests side-by-side, and use golden sets as baselines. This applies to teams building on the Google Cloud Developer plugin or similar agentic stacks.
What to check- Implement trajectory-based evaluation suites rather than output-only checks.
- Preserve turn structure in telemetry to debug parallel dispatch issues.
- Run parallel live and plan-only evaluations to isolate reasoning from permission errors.
Governance areas it touches: Tool / action control · Identity and zero trust · Observability and audit
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16
ANNOUNCED
Google Research
· 2026-10-16
Google describes a reference architecture separating cognitive orchestration from GPU execution. It uses Agent Runtime, GKE, and native governance features like Agent Identity and Model Armor to manage high-compute biotech agents with asymmetric human oversight.
What to checkImplement asymmetric gating to require human approval for irreversible actions. Use cryptographic agent identities to ensure auditability of specific agent decisions. Enforce data residency by routing proprietary data through internal VPC endpoints rather than external proxies.
Governance areas it touches: Tool / action control · Human-in-the-loop support · Identity and zero trust · Observability and audit · Data protection · Inter-agent communication security
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16
ANNOUNCED
Microsoft Feature
· 2026-10-13
Microsoft announced Azure Content Understanding 2.0 Preview prebuilt analyzers for invoices, tax forms, and identity documents. These tools provide structured extraction with confidence scores and source grounding to support human review and validation workflows.
What to check- Configure confidence thresholds to trigger mandatory human review for low-score extractions.
- Validate that grounding evidence is logged for audit trails.
- Test prebuilt analyzers against representative edge cases before production deployment.
Not addressedTool / action control
DirectlyHuman-in-the-loop support
Not addressedIdentity and zero trust
PartlyObservability and audit
Not addressedData protection
Not addressedInter-agent communication security
Why these scores
- Tool / action control: The document describes extraction capabilities but does not detail tool allowlisting or sandboxing controls.
- Human-in-the-loop support: The source explicitly states confidence scores can inform human-review thresholds and route uncertain results to reviewers.
- Identity and zero trust: No specific agent identity, authentication, or least-privilege mechanisms are described in the text.
- Observability and audit: Grounding provides source evidence for results, but comprehensive tracing or audit logging is not detailed.
- Data protection: The text mentions PII in examples but does not describe residency, retention, or protection controls.
- Inter-agent communication security: The item focuses on single-agent document processing, not secure communication between agents.
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16
ANNOUNCED
Microsoft Feature
· 2026-10-14
Microsoft announced Azure Content Understanding features that link extracted values to source evidence and provide confidence scores. This enables human-in-the-loop routing based on risk, reducing review burden by focusing attention on low-confidence fields.
What to check- Calibrate confidence thresholds per document type and field using representative production data.
- Implement routing rules that direct low-confidence extractions to human review.
- Maintain audit trails linking extracted values to their source document evidence.
Not addressedTool / action control
DirectlyHuman-in-the-loop support
Not addressedIdentity and zero trust
DirectlyObservability and audit
PartlyData protection
Not addressedInter-agent communication security
Why these scores
- Tool / action control: The document focuses on document extraction accuracy and does not describe tool allowlisting or sandboxing controls.
- Human-in-the-loop support: The source explicitly describes using confidence scores to route results to human review and calibrate thresholds for HITL workflows.
- Identity and zero trust: No mention of agent identity, authentication, or least privilege mechanisms is present in the text.
- Observability and audit: The text states that grounding allows keeping an audit trail from the business record back to the page it came from.
- Data protection: The document notes that customer data is not used to train the confidence model, but does not detail residency or retention policies.
- Inter-agent communication security: The workflow described is single-agent document processing, with no discussion of inter-agent communication security.
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16
ANNOUNCED
Microsoft Release
· 2026-10-15
Microsoft Foundry announced GA for Routines, Tool Search, and Agent-to-Agent (A2A) protocols. Insights entered public preview for trace-backed investigation. Model Router added preview session affinity and routing metadata. GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna are generally available.
What to check- Review A2A authentication mechanisms and ensure Invocations use scoped credentials.
- Validate Insights trace data handling to ensure PII is excluded from telemetry.
- Test Tool Search allowlisting to prevent unauthorized tool discovery.
DirectlyTool / action control
PartlyHuman-in-the-loop support
PartlyIdentity and zero trust
DirectlyObservability and audit
PartlyData protection
DirectlyInter-agent communication security
Why these scores
- Tool / action control: Tool Search is GA and loads tool definitions on demand, directly addressing tool management.
- Human-in-the-loop support: The workflow example emphasizes reviewing notes and asking before deploying, but this is procedural guidance rather than a built-in control.
- Identity and zero trust: The example uses authenticated Invocations, but the announcement does not detail specific identity or least-privilege controls for agents.
- Observability and audit: Insights enters public preview to investigate trace-backed findings, and Model Router adds routing metadata for inspection.
- Data protection: The example advises keeping secrets out of telemetry and content capture off by default, but this is a configuration recommendation, not a new platform feature.
- Inter-agent communication security: Agent-to-Agent (A2A) is GA, and the example describes using authenticated Invocations for handoffs between agents.
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16
SHIPPED
OpenAI Release
· 2026-10-14
OpenAI Agents SDK v0.24.0 exposes MCP annotations on converted tools and supports run-wide tool exception formatting. It also defers automatic compaction for pending approvals and batches conversation session writes within API limits.
What to checkReview custom tool error formatters to handle the new 'tool_exception' category. Treat exposed MCP annotations as untrusted metadata and do not rely on them for authorization decisions.
DirectlyTool / action control
PartlyHuman-in-the-loop support
Not addressedIdentity and zero trust
PartlyObservability and audit
Not addressedData protection
Not addressedInter-agent communication security
Why these scores
- Tool / action control: The release explicitly adds support for run-wide tool exception formatting and exposes MCP annotations on converted tools.
- Human-in-the-loop support: The release defers automatic compaction for pending approvals, indicating support for approval workflows, but does not describe new interactive controls.
- Identity and zero trust: The document does not describe changes to agent identity, authentication, or least privilege mechanisms.
- Observability and audit: The release includes fixes for tracing opt-out and lifecycle debug logs, but does not introduce new audit features.
- Data protection: The document does not address PII, data residency, or retention policies.
- Inter-agent communication security: The document does not describe security controls for communication between agents.
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16
SHIPPED
OpenAI Release
· 2026-10-16
OpenAI Python SDK v3.33.0 adds safety identifiers to image requests and responses. It also shares the Decisions cache breakpoint parameter type. These changes are included in the versioned release.
What to check- Update SDK to v3.33.0 to capture safety identifiers.
- Log these identifiers in audit systems for image generation compliance.
Not addressedTool / action control
Not addressedHuman-in-the-loop support
Not addressedIdentity and zero trust
PartlyObservability and audit
Not addressedData protection
Not addressedInter-agent communication security
Why these scores
- Tool / action control: The release notes do not mention tool allowlisting, permissions, or sandboxing controls.
- Human-in-the-loop support: No human-in-the-loop approval or review mechanisms are described in the changelog.
- Identity and zero trust: The update does not address agent identity, authentication, or least privilege enforcement.
- Observability and audit: Adding safety identifiers to image requests and responses provides data points for potential audit trails.
- Data protection: The changelog does not specify PII handling, data residency, or retention changes.
- Inter-agent communication security: No changes to inter-agent communication security are documented in this release.
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16
SHIPPED
Anthropic Release
· 2026-10-15
The release adds a speed parameter to the Messages API and documents headers for Managed Agents and telemetry endpoints. It also covers OpenTelemetry options in tests. These changes support observability and configuration of response characteristics.
What to check- Integrate the new telemetry endpoints into your monitoring stack.
- Verify OpenTelemetry configuration to ensure agent actions are logged.
- Review the speed parameter to manage latency and cost implications.
Not addressedTool / action control
Not addressedHuman-in-the-loop support
Not addressedIdentity and zero trust
DirectlyObservability and audit
Not addressedData protection
Not addressedInter-agent communication security
Why these scores
- Tool / action control: The changelog does not mention tool allowlisting, sandboxing, or permission controls.
- Human-in-the-loop support: No features related to human approval or intervention are described in the changelog.
- Identity and zero trust: The release does not address agent identity, authentication, or least privilege mechanisms.
- Observability and audit: The source explicitly mentions documenting telemetry endpoints and covering OpenTelemetry options in tests.
- Data protection: No changes regarding PII handling, data residency, or retention policies are listed.
- Inter-agent communication security: The release notes do not describe security controls for communication between agents.
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16
ANNOUNCED
Anthropic Policy
· 2026-10-14
Anthropic announced enhanced election safeguards, including new classifiers for multi-step misuse and integration of Democracy Works voting data. The company reported that Claude Opus 5.5, Sonnet 5.5, and Haiku 5.5 scored 99% on political evenhandedness evaluations, while detailing threat intelligence findings on coordinated abuse.
What to check- Review vendor-provided election and bias evaluation reports to verify alignment with internal compliance standards.
- Monitor vendor threat intelligence updates for emerging multi-step misuse patterns.
- Ensure internal logging captures agent interactions to support independent audit requirements.
Governance areas it touches: Tool / action control · Identity and zero trust · Observability and audit · Data protection
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16
ANNOUNCED
Anthropic Policy
· 2026-10-16
Anthropic announced that METR has received employee-like access to internal tools and systems for risk assessment. This follows a commitment to independent third-party evaluation under the White House Accord. Other partners include SecureBio and Accenture, with a three-tier access framework defined for future evaluators.
What to check- Establish independent third-party audit channels for AI safety.
- Define clear data redaction categories for external evaluators.
- Implement tiered access levels for external security assessments.
Governance areas it touches: Human-in-the-loop support · Identity and zero trust · Observability and audit · Data protection
Analysed by Qwen 3.8 (27B) · Source checked 2026-10-16