Standalone-api / brainstore v2.1.1 ships starlette < 1.0.1 (CVE-2026-48710) — please rebuild with starlette >= 1.0.1

[joel-vaz]

Summary

The current self-hosted data-plane images
public.ecr.aws/braintrust/standalone-api:v2.1.1 and
public.ecr.aws/braintrust/brainstore:v2.1.1 (pinned by Helm chart 6.2.2,
the latest as of 2026-05-28) include starlette < 1.0.1 in their
dependency tree, which is affected by CVE-2026-48710 / GHSA-86qp-5c8j-p5mr
(CVSS 6.5, "BadHost" Host-header path poisoning).

starlette 1.0.1 (the fix) was released 2026-05-21, but the v2.1.1 images
were built 2026-05-08 — before the fix — so they still carry the vulnerable
version. No newer image tag or chart release currently addresses it.

Vulnerability

starlette < 1.0.1 does not validate the HTTP Host header before constructing
request.url. A Host header containing a path component poisons
request.url.path while the ASGI-routed request.scope["path"] is unchanged,
allowing middleware that relies on request.url.path for auth/routing to be
bypassed.

Advisory: GHSA-86qp-5c8j-p5mr
Fix release: https://pypi.org/project/starlette/1.0.1/

Affected

• Image: public.ecr.aws/braintrust/standalone-api (and brainstore)
• Tag: v2.1.1 (built 2026-05-08), pinned by chart braintrust 6.2.2
• starlette versions observed in deps: 0.49.3, 1.0.0

Request

Please publish a rebuilt data-plane image (and a chart release that pins it)
with starlette >= 1.0.1. We're tracking this internally and would
appreciate an ETA so we can schedule the upgrade.

Verification we'll run on the new image

docker run --rm public.ecr.aws/braintrust/standalone-api: pip show starlette

expect: Version: 1.0.1 (or higher)

A couple of judgment calls worth flagging:

• I kept the issue non-exploit-detail and fix-focused — appropriate for a public issue on a known/published CVE.
• If you'd rather not open it publicly at all, the email-only route is fine; the chart repo issue just tends to get faster maintainer eyes.