← All updates

Security / News explained

Braket’s security updates cover inputs and results

AWS’s November 9 bulletin fixes two different paths from writable S3 data into a local machine: input downloads in the Braket SDK and result validation in Braket Schemas. Both package versions and the writer’s permissions matter.

By Zyel · Published · 4 min read

AWS published two Amazon Braket security fixes on November 9, 2026. They address different transitions: downloading input data for a local hybrid job and validating a quantum task’s result. In each case, a remote authenticated user needs write access to the relevant S3 location. [1]

The Braket SDK is the Python library used to run quantum tasks and hybrid jobs and retrieve results; Braket Schemas helps parse those results. The bulletin identifies separate corrected versions for these two packages. [1]

Zyel’s interpretation is that the important boundary is the machine consuming the data. Cloud storage is part of the workflow, but the authority involved in a download or a parser operation belongs to the process that performs it. Reviewing only where the quantum task ran would leave that part of the system out of the picture.

Input names could escape the download directory

CVE-2026-124034 affects amazon-braket-sdk versions from 1.10.0 up to, but excluding, 1.127.7; version 1.127.7 fixes it. Crafted S3 object keys could make a local hybrid job overwrite files outside its intended input directory under the job user’s privileges. [2]

The project advisory confines this issue to local hybrid jobs and says jobs run by the Amazon Braket service are unaffected. It describes file writing, rather than file reading or direct code execution. [2]

For a hypothetical research team, the local job might run in a workstation environment used to prepare an experiment. If another identity can supply objects to that job’s input location, the review needs to account for that writer as well as the person starting the job. The intended destination directory is part of the security promise the download should keep.

The distinction between a file-write flaw and direct execution is useful for response planning. Our view is that a team should describe the documented capability accurately, then assess its consequences in the actual environment. Broadening every file-write advisory into an unconditional execution claim makes that assessment less precise.

A result document could select Python behavior

CVE-2026-124035 affects amazon-braket-schemas from 1.17.0 up to, but excluding, 1.32.2; version 1.32.2 fixes it. The advisory says error-mitigation validation could import and call Python code named inside a task result document. [3]

Exploitation requires write access to the victim’s task-output S3 location and the victim retrieving the affected result. Code would run with the result-retrieval user’s privileges on the machine retrieving it. [3]

That is a different trust decision from saving an input file. A result can be valid-looking data at the storage boundary while still influencing behavior when a library interprets it. In our reading, the consumer’s software version and authority therefore deserve the same attention as the permissions around the output object.

Imagine a hypothetical analyst who completes a task in one environment and opens its result later in another. A review limited to the original submission environment could miss the second consumer. This is why a useful inventory follows the data through the workflow: which environments prepare local jobs, and which environments subsequently retrieve and interpret results?

Two packages require a complete dependency decision

AWS calls the Schemas change a follow-on fix to CVE-2026-9291. It recommends incorporating the fixes into derivative code and restricting write access to job-input and task-output locations, including the job’s role, to trusted principals. [1]

The practical implication is to verify the resolved versions in each relevant environment. A request to update “the SDK” is a starting instruction, but completion evidence should show what happened to both affected packages. A dependency declaration, a prepared environment and the process that actually consumes the data are distinct points where that answer can be checked.

For a hypothetical shared workflow, the owner could record the corrected package pair, identify the intended writers for each data location and confirm an approved ordinary input-and-result cycle. Those are proposed acceptance criteria, not claims that Zyel tested a customer’s quantum workload.

The broader lesson is about authority following data. Permission to contribute an input or a result should remain a bounded contribution to that workflow. The local file system and the result consumer’s Python process need their own protections when that contribution is downloaded and interpreted.

Sources + further reading

Source dates describe the original material. Publication above is the date this Zyel article went live.

  1. AWS bulletin 2026-175: Amazon Braket SDK and SchemasAWS Security · Checked Nov 10, 2026
  2. Braket SDK local hybrid job input download advisoryAmazon Braket project · Checked Nov 10, 2026
  3. Braket Schemas error mitigation validation advisoryAmazon Braket project · Checked Nov 10, 2026

Reader discussion

What would you ask next?

Questions, useful corrections and thoughtful disagreement are welcome. Comments are reviewed before they appear.

No reader comments yet.

10–2,000 characters. Plain text. Leave out passwords, private project data and personal contact details. A blank display name appears as Anonymous.

Complete the CAPTCHA before sending. Verification runs on your device.