Disclosed weakness · no CVE recorded

LiteLLM keyless JSON providers could receive the OpenAI key

LiteLLM · LiteLLM · JSON-configured providers

First documented disclosure · Review checkpoint

Summary

A provider deployment without its own credential could send the process’s OpenAI key to another provider or a configured endpoint.

Severity and exploitation evidence

No CVSS score recorded in this review.

Exploitation not established: Reviewed maintainer reports and controlled checks do not establish malicious exploitation against users.

Exploitation evidence source

Technical severity is not a measure of observed attacks or of risk to your deployment. Attempts, demonstrations, and successful compromise retain the qualifications in the source.

How the weakness works

OpenAI-compatible request paths fell back to OPENAI_API_KEY instead of rejecting a JSON-configured provider with no provider-specific credential.

Conditions required

A JSON-configured provider lacks its own key while an OpenAI key is available to the process; the requested endpoint receives the resulting authentication header.

Potential impact

Disclosure of the OpenAI credential to the selected upstream. Controlled checks used a fake OpenAI key; unauthorized use of a real credential was not established.

Affected versions and configurations

The sources demonstrate affected pre-fix builds but do not establish the complete package-version range.

Fix and mitigation status

Stable 1.108.15 includes backport PR53314; the tagged shared guard and call sites were reviewed. Development 1.112.0-dev.2 also includes PR50819.

Practical next steps

Upgrade the affected LTS deployment, set the intended provider credential and rotate an OpenAI key if affected calls reached an untrusted destination.

Operating lesson

A missing provider credential should fail locally rather than borrow another provider’s secret.

Evidence and limits

The feature PR opened November 10; the specific credential-fallback correction is documented by the November 17 public review timeline and patch timestamps. The stable release is November 21. Explicit credentials, caller-supplied OpenAI clients and deliberately forwarded Authorization headers remain supported; the correction is not a blanket ban on custom endpoints. This is distinct from the existing OpenAI SDK 3.x administrator-key fallback record, which tracks PR50218.

Sources