Defaults that decide: GPUStack's loopback admin, dated Compute Engine APIs, spend caps
Three stories this week come down to a single default value, and to what a project does when that default turns out to be wrong.
GPUStack's was force_auth_localhost=false. With that setting, any request the software considered to come from loopback was promoted to the built-in administrator account. The GitHub advisory for CVE-2026-102895 explains why the default failed in practice: the common way to run GPUStack is behind a reverse proxy on the same host, which makes every remote request arrive from 127.0.0.1, and a forged X-GPUStack-Real-IP header achieves the same result without any proxy. SANS reported exploit attempts on 12 November. The fix in version 2.2.0 did not flip the default; it removed the loopback auto-admin path altogether.
Google took the opposite route with Compute Engine. Its /v1 API had defaults that the team itself calls counter-intuitive, such as returning no partial results from AggregatedList unless a flag is set, but changing them on a surface shared by every customer would break scripts. The answer, according to the post InfoQ reported on 14 November, is to leave /v1 frozen with no end-of-life date and put the better defaults, partial results, structured errors and tighter IAM checks, into dated versions that callers choose explicitly or receive through new client libraries.
Cloudflare's change is about a default nobody chose: unlimited spending. Its existing alerts only informed; the new Spend Management beta lets a customer set a limit at which Workers or Durable Objects stop. The post is careful about the new default's edges. Only billable customer code is paused, static assets keep serving, and spend is evaluated with a delay of up to about ten minutes, so the cap is a brake rather than a wall.
Put side by side, the three show the available moves. A default that grants trust can only be removed, because no migration path makes it safe. A default that shapes behaviour can be versioned, so old callers keep it and new callers opt out. And a default that leaves risk open can be closed with a new control, as long as its side effects are documented.

Why it matters
Read together, the three cases suggest a review question for any service is which of its defaults grants trust or removes a limit, and whether the project has a way to change it that does not depend on every user reading release notes.