AWS AI Updates: November 11, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. OpenAI’s Daybreak Cybersecurity Models, GPT-5.6 Sol and GPT-5.6 Cyber, Reach Eligible Bedrock Customers on the bedrock-runtime Endpoint
Amazon Bedrock added support on November 10 for two OpenAI cybersecurity models, Daybreak Blue (GPT-5.6 Sol) and Daybreak Red (GPT-5.6 Cyber), on the bedrock-runtime endpoint through US and Global cross-Region inference profiles. AWS positions Daybreak Blue as the starting point for most security teams, covering vulnerability discovery, detection engineering and incident response. Daybreak Red targets advanced authorized work such as vulnerability research, exploit reproduction and mitigation development, with a lower refusal threshold offset by stronger identity verification, monitoring and access controls. Access is limited to eligible customers: they enroll through OpenAI, then work with their AWS account team to request access on AWS. AWS states that inference data is not used for model training and that neither model requires opting in to data sharing with OpenAI. The announcement does not give pricing, quotas or context size, and refers to the documentation for bedrock-mantle endpoint Regions. Source
2. ElastiCache Serverless for Valkey 9 Adds Vector, Full-Text and Hybrid Search, Targeting Agent Memory and RAG
Amazon ElastiCache Serverless now supports vector search, full-text search, hybrid search and server-side aggregations on live data, announced November 10. AWS cites latency as low as microseconds at up to 99% recall for vector search, with updates searchable as soon as writes complete. Search requires ElastiCache version 9 for Valkey, and existing serverless caches can be upgraded or new ones created. AWS names agent memory and context retrieval (a separate datastore per agent, session or task), retrieval-augmented generation, and product discovery and recommendation as use cases. The feature is available in all commercial Regions, the AWS GovCloud (US) Regions and the AWS European Sovereign Cloud Region, billed on usage. The announcement does not describe index types, distance metrics, query syntax or hybrid ranking. Source
3. AWS Publishes a Playbook for Choosing Built-In, DeepEval and AutoEval Evaluators in AgentCore Evaluations
Amazon Bedrock AgentCore Evaluations is the subject of an AWS Machine Learning Blog post dated November 10 on how to evaluate production agents. The post recommends defining success in business terms, breaking it into observable behaviors, and mapping each behavior to a trace-level or session-level evaluator. Built-in evaluators are referenced by ID (for example Builtin.GoalSuccessRate) and cover response quality, task completion, tool selection and parameter accuracy, safety, and trajectory matching. Third-party managed evaluators from DeepEval (such as TaskCompletion, PIILeakage and ConversationCompleteness) and AutoEval (Security, Possible, Humor) run on AgentCore-hosted code with AgentCore selecting the judge model. Custom LLM-as-a-judge evaluators can score several named dimensions in one call and return one result per dimension with no composite score, so releases can be gated per criterion. Evaluators that need ground truth, such as TrajectoryAnyOrderMatch and Correctness, run only in batch evaluation because live traffic has no reference answer. On-demand evaluation uses client.evaluate() on the bedrock-agentcore boto3 client with an evaluatorId and sessionSpans. Source
4. AWS Sample Builds a Cross-Account Data Lake of Bedrock Invocation Logs for Token Cost Attribution
Amazon Bedrock invocation logs are the input to a sample “tokenomics” data lake that AWS described on November 10. Bedrock writes model invocation logs to CloudWatch in each source account, subscription filters forward them to a central account, a Lambda function unwraps and flattens the batches, and Amazon Data Firehose converts them to Parquet in S3 for querying with Glue and Athena. The resulting table includes the prompt, response, input, output and cache token counts, model, stop reason (such as max_tokens), account, timestamp and IAM identity, and it can be joined with Cost and Usage Report 2.0 to attribute spend by team or account. An optional Amazon Quick Sight dashboard deploys through the Cloud Intelligence Dashboards framework. Deployment uses two CloudFormation templates, with the code in the aws-samples/sample-bedrock-token-cost-allocation repository, and enabling invocation logging remains a separate account-level step. The post gives no cost or benchmark figures. Source
5. TCS Reports 88% Less Effort From an Agentic Test Orchestrator Built on AgentCore and Strands Agents
Tata Consultancy Services described on the AWS Machine Learning Blog (November 10) an Agentic Test Orchestrator, a multi-tenant SaaS platform for banking, financial services and insurance testing. It ingests Jira stories, business requirement documents, Swagger specifications and captured UI flows, then generates scenarios, Gherkin feature files and executable Java or Python tests and runs them. Five agents built with the Strands Agents SDK run on the Amazon Bedrock AgentCore runtime, each session in its own microVM, with a deterministic supervisor routing work, Bedrock Knowledge Bases filtered by tenant and project, and tests executed on ECS (Fargate for API tests, EC2-backed Selenium for UI tests). The execution agent is a LangGraph state graph with a bounded retry loop, and a fixing agent handles only code errors, not environment errors. On an internal benchmark of about 100 scenarios, TCS reports about 2 hours per scenario manually versus about 14 minutes with the platform, roughly 200 hours versus 24 hours in total, an effort reduction of about 88%. The post states results vary by workload and does not name the models used. Source
6. Trovata Moves Its Treasury Platform From AI Insights to Scheduled Agents on Bedrock and AgentCore
Trovata, a cash and liquidity management platform, described on the AWS Machine Learning Blog (November 10) how it runs autonomous treasury agents on Amazon Bedrock and Bedrock AgentCore. Agents execute recurring work on schedules, triggers or rules, including cash forecasting, anomaly detection, and reconciliation summaries across hundreds of accounts, and are configured through an Agent Builder-Configurator console. Bedrock supplies LLM inference, AgentCore supplies orchestration and an isolated runtime, and agents operate under the same role-based access controls as human users. The AI retrieves data only through Trovata’s verified API layer with no direct database access, and queries and agent actions are logged for compliance review. Trovata reports annualized savings above $1M, more than 5 hours saved per day per accounting team and a 77% annual reduction in treasury management system cost, figures drawn from its customer case studies and not independently verified. The post does not name the foundation models used. Source