Friday, October 23, 2026
AWS made Continuum for code vulnerabilities generally available, gave CloudFormation a one-step force rollback, opened SageMaker AI training session APIs, and previewed Interconnect last mile with Megaport. OpenSearch Serverless now refreshes in 5 seconds and Connect agents get tailored Customer Profiles. Azure shipped vectored reads for Data Lake Storage, and OpenAI made its Decisions API generally available. Theme: the quick path is fine, as long as someone owns the production call, same as today's Free Jack on SageMaker JumpStart.
AWS: Continuum for code vulnerabilities is generally available
AWS Continuum for code vulnerabilities builds a context graph of your AWS environment, including resource topology, identity relationships, and network reachability. It ingests your existing vulnerability findings, scans your code, confirms which findings are actually exploitable by testing them in an isolated sandbox, ranks them by business impact, and then proposes fixes.
A scanner that proves the finding is real before paging anyone. The security backlog may finally shrink below the height of the CN Tower.
AWS: CloudFormation adds force rollback for stuck stack updates
When some resources cannot roll back automatically during an update rollback, CloudFormation can now force the rollback to finish in a single operation and return the stack to a deployable state. Previously you had to identify, list, and retry each affected resource. It works from the CLI, SDKs, AWS CDK, and AI agents, and any resource that still needs attention is flagged for follow-up.
One button to get back to a known good state. The resource it skipped is still sitting there, waiting for a volunteer.
AWS: SageMaker AI adds training session APIs for serverless customization
SageMaker AI serverless model customization now offers training session APIs, so teams can write their own post-training loop without provisioning GPU infrastructure. Operations such as forward_backward(), optim_step(), and sample() become API calls, which allows custom loss functions, reward signals, or data curricula alongside the managed SFT and DPO techniques.
Fits neatly with today's Free Jack on JumpStart. The prototype is easy. Owning the training loop is the production decision.
AWS: Interconnect last mile enters preview with Megaport
AWS Interconnect last mile is a managed connectivity preview, built with Megaport, that connects branch offices, data centers, and remote locations to AWS from the console by choosing a Region and bandwidth. Separately, Direct Connect added 100 Gbps dedicated connections at its Netmagic DC2 location near Bangalore.
Private network to AWS in a few clicks. The procurement form for the circuit is sold separately.
AWS: OpenSearch Serverless now refreshes every 5 seconds by default
Amazon OpenSearch Serverless cut the default refresh interval from 10 seconds to 5 seconds for search and vector collections on its next generation, so newly ingested documents become searchable sooner. No configuration change is needed.
Half the wait for read-after-write. Your users will now notice the other delays instead.
AWS: Connect agents get custom Customer Profiles layouts
Amazon Connect administrators can now build Customer Profiles layouts in Profile Explorer, choosing which widgets such as contact history, contact information, and lifetime value attributes each agent sees in the agent workspace.
Agents see what they need for the call in front of them. A rare case of less on screen being the feature.
Azure: ABFS Vectored Read for Data Lake Storage is generally available
ABFS Vectored Read lets compatible analytics engines combine nearby file ranges into fewer storage requests and run reads asynchronously, which Microsoft says can reduce network round trips for selective reads in Azure Data Lake Storage.
Fewer requests for the same data. Your transaction line on the bill may quietly approve.
AI: OpenAI Decisions API is generally available
The Decisions API, built on GPT-6 Luna, answers bounded questions (a probability, a choice from your options, or a score against a rubric) from text, images, or JSON. OpenAI says it is up to 10 times faster than Luna through the Responses API, allows up to 400 questions per request, offers US and EU inference residency, and costs $0.10 per million input tokens with no output-token charges.
Classification, routing, and next-step calls priced like a rounding error. Somebody still owns what happens when it picks wrong.
AI: OpenAI partners with Poland's NASK on AI safety
OpenAI and NASK, Poland's national research institute, announced a partnership on AI safety, security, and technical research aimed at strengthening safeguards.
Another safety partnership on the board. The fine print is where the actual work lives, as usual.