AWS Gave Kubernetes Pods Their Own VM. What Can the Worker Still Write?

A supplier sends a price sheet. A worker reads it and proposes a new price for a case of tomatoes. If that same worker can overwrite the price buyers see, putting it in its own virtual machine still leaves it with the authority to change a quote. AI generated this article from public sources. The worker design and staging exercises are illustrative. On November 9, AWS launched Nitro Sandboxes for Amazon EKS, ECS, and EC2. On EKS, an opted-in pod runs in its own hardware-isolated virtual machine on a supported EC2 node. AWS documents that a networked sandbox can still use EKS Pod Identity to obtain temporary credentials for the IAM role associated with its service account. That makes a useful first question for a rollout. If the parser misbehaves, what would its credentials still let it do? AWS launch announcement · EKS Pod Identity documentation Give extraction a place to stop The delivered TMC Price Intelligence engagement turned supplier emails and attachments into comparable records. The public case study describes extraction, a central pricing repository, reviewable reports, and quote and invoice workflows. The proposed architecture below uses that separation as a starting point. It’s an example for a new intake pipeline. TMC Price Intelligence case study Imagine two storage areas. An intake service puts an attachment in the first and assigns a job ID. A parser receives that job’s input reference and writes a candidate to the second. The candidate has the item, pack size, amount, and a reference to where the parser found each value. The worker identity can read the assigned input and write its candidate. Scope those permissions using the workload’s actual isolation needs. A shared role with access to every input in a bucket still has that access inside a sandbox. A separate service takes the candidate through validation and review. It resolves the supplier and job from records the intake service created, since both values could be wrong in the candidate. It can send an unfamiliar pack size to review and record which revision was approved. The quote service reads the approved record. That gives the system a place to stop when a source file is valid enough to parse but too ambiguous to price from. Rehearse the permissions in staging Start with a synthetic spreadsheet and a worker running under the same kind of service account you plan to ship. Have it read its assigned object and write one candidate. Then have the same identity attempt a write to a dummy approved-price record through the same endpoint the application uses. Expect a denial even if the request contains a plausible price and valid job ID. Read the record back so the test checks the outcome as well as the response. If inputs are meant to be separated by job or tenant, include a forbidden read against a second synthetic input. Next, submit a valid-looking candidate that points to a different job or lists a pallet price as a case price. Check that it stays out of the approved records and that the reviewer can see the original value. This exercises a different boundary. The parser may be authorized to produce a candidate, and a candidate can still be wrong. When a buyer receives a quote, both checks matter. Nitro Sandboxes can also remove a pod’s network access. AWS says Pod Identity credentials are unavailable in that configuration. A worker designed to fetch inputs and write candidates directly to cloud services needs a network path and narrowly scoped permissions. A networkless variant needs a different way to deliver input and collect output. Test whichever path you deploy with the same allowed and forbidden operations. EKS sandboxing guide · EKS credentials documentation The extra stage costs a little engineering. Someone has to operate the candidate store, handle ambiguous units, and recover failed imports. It also gives operations a smaller decision to make when something goes wrong. They can quarantine a set of candidate prices, inspect the source sheets, and keep quoting from the last approved records while they sort it out.