Daily News · 3 min read

AWS AI Updates: October 23, 2026

aws sagemaker bedrock agentcore fine-tuning reinforcement-learning open-weight-models ai-agents security continuum

1. SageMaker AI Training Session APIs Let Teams Write Their Own Post-Training Loops on Serverless GPUs

AWS. Amazon SageMaker AI serverless model customization now offers training session APIs, which expose operations such as forward_backward(), optim_step(), and sample() as calls against a live, stateful training session. Teams can write their own training loop (custom losses, new RL reward signals, data curricula, or algorithms from recent research) while SageMaker AI handles GPU allocation, model and LoRA loading, distributed execution, fault recovery, checkpointing, and security. Billing is per token processed, so local computation such as reward scoring is not billed as GPU time, and AWS is publishing a SageMaker training sessions cookbook with ready-to-run SFT and RL recipes. The APIs are generally available in US East (N. Virginia), US West (Oregon), and Asia Pacific (Mumbai). Source

2. AgentCore Gateway Adds Bedrock InvokeModel Support and Automatic Model Fallback

AWS. Amazon Bedrock AgentCore Gateway’s inference endpoint now accepts the Bedrock InvokeModel and InvokeModelWithResponseStream APIs, so existing Bedrock SDK integrations can be pointed at the gateway to reach any model provider through one governed endpoint. A new model fallback feature routes requests to an ordered chain of alternative models or providers when the primary destination is throttled, out of capacity, or unreachable, with trigger conditions such as throttling, server errors, or response timeouts defined in a gateway rule. A built-in circuit breaker skips destinations that are currently failing, and the client sends one request and receives one response regardless of internal retries. Both features are available in all Regions where AgentCore Gateway is offered, except China. Source

3. AWS Continuum for Code Vulnerabilities Reaches General Availability

AWS. AWS Continuum for code vulnerabilities, which was in gated preview after Continuum’s June announcement, is now generally available. The service builds a context graph of an AWS environment (resource topology, identity relationships, network reachability), ingests existing vulnerability findings, scans code, and confirms exploitability by testing findings in an isolated sandbox before prioritizing them by business impact. It delivers remediation patches as pull requests with exploitability proof and a blast radius assessment attached, and selects an AI model per task. Customers start in a Learn mode where they approve proposed fixes, then can switch to an Enforce mode where Continuum acts autonomously within configured guardrails, alongside services such as Amazon GuardDuty and AWS Security Hub. Source

4. SageMaker Serverless Customization Adds Muse Glimmer, Qwen3.8 27B, and Ministral 3

AWS. SageMaker AI serverless model customization now supports Muse Glimmer, Qwen3.8 27B, and Ministral 3 in 3B, 8B, and 14B sizes. The models can be adapted with supervised fine-tuning, direct preference optimization, or reinforcement fine-tuning, using either LoRA or full fine-tuning. They join a lineup that already includes the Nova, Gemma 4, Nemotron 3.5, Qwen, Llama, gpt-oss, and DeepSeek families, available in US East (N. Virginia), US West (Oregon), Asia Pacific (Tokyo), and Europe (Ireland). Source