AWS Continuum tests exploits in a sandbox, then opens the fix as a pull request

Amazon Web Services made AWS Continuum, a vulnerability management service for code, generally available in its October 26 weekly roundup. It proves findings in a sandbox and opens the fix as a pull request.

A scanner that proves the bug and proposes the patch

According to AWS, Continuum treats vulnerability management as an end-to-end job instead of a scanner that returns a list. It builds a context graph from a customer’s repositories and AWS environment and gives each finding an evidence level. Exploitable findings are reproduced in an isolated sandbox with proof-of-concept code, so a reviewer sees the issue itself and not just a severity label.

Issues are then ranked by how close the attack path runs to business-critical systems, and the fix is opened as a pull request. For now every remediation needs human approval: the customer reviews and merges. AWS says autonomous fixes for some classes of problem may come later, gated behind confidence thresholds.

Continuum is available in US East (N. Virginia) and US West (Oregon). The post gives no pricing, so cost is not established.

Honeypot data behind the timing argument

AWS tied the launch to its own telemetry. Amazon Threat Intelligence says its MadPot honeypots recorded a PHP2Shell exploitation attempt on October 17, four days before CVE-2026-109618 and CVE-2026-109619 were disclosed. On disclosure day, proof-of-concept exploits were reportedly circulating around 02:00 UTC, about twelve hours before the 14:00 UTC embargo end.

The author draws a blunt conclusion: “a patch window measured in days assumes a delay that no longer exists”. AWS also published AWS WAF mitigations for the same CVEs. These are AWS’s own observations and the post does not link independent confirmation.

Other launches in the roundup

• Amazon ECS Gateway lets platform teams define one gateway with subnets, certificates, listeners and AWS WAF rules, while application teams opt in through a gateway connector. Up to 25 services share one Application Load Balancer. It is available wherever ECS runs, at no extra charge.
• Amazon Polly upgraded its Generative engine to 57 voices across 11 languages and 25 locales. It adds 14 voices, including Brazilian Portuguese voices Leo and Carolina, with the same SynthesizeSpeech call and the same price.
• Structured outputs for Claude on Amazon Bedrock return schema-compliant JSON from any Anthropic Claude model, on both the bedrock-runtime and bedrock-mantle endpoints.
• SageMaker AI training sessions are generally available. Customers control custom loss functions and reward signals, and AWS handles GPU allocation and recovery. Billing is per token processed.
• ElastiCache Serverless for Valkey supports online migration from self-managed Valkey or Redis OSS on EC2 with, AWS says, zero downtime.

TorchNeuron lowers the cost of trying Trainium

The most consequential item for AI teams may be TorchNeuron, an open source PyTorch backend in public beta. Per AWS, a model moves to Trainium by switching to .to('neuron'), keeping eager execution and torch.compile. It works on Trainium2 and Trainium3 and integrates with the Neuron Kernel Interface for custom operators. It arrives alongside AWS Neuron 2.33.

The hard part of non-GPU hardware has usually been software: models written for one stack need porting. A native PyTorch path removes much of that, because custom operators remain the only place where deeper work is needed.

Early outside evidence is limited. In our report on a Hugging Face Space that tracks Trainium support, Hugging Face said 89% of the 100 most-used text-generation architectures work out of the box, against 46% with Optimum Neuron v0.4.5. Those are the company’s figures and were not independently verified. That testing covers only Trainium2, text-generation and image-text-to-text tasks, and models that fit on four devices.

The demand side is visible too: Verizon’s more than $2 billion deal for distributed AI compute, covered in our earlier article, shows how much money is flowing into AWS infrastructure.

What the roundup leaves unproven

The post is a summary, so benchmarks, pricing for Continuum and detail on how evidence levels are scored are absent. Continuum’s value depends on how often its sandbox reproductions are correct and how many generated pull requests teams merge, and neither figure is published. The roundup also mentions Agent Registry for discovering MCP servers and agent skills at runtime, and a cost study of coding agents that found routing 44.7 percent cheaper for 4.8 percent lower mean quality, again AWS’s own results.

Looking ahead, AWS re:Invent runs November 30 to December 4 in Las Vegas.