Disclosed weakness · no CVE recorded

LiteLLM forged OAuth headers could suppress recorded spend

LiteLLM · LiteLLM Proxy · Anthropic subscription accounting

First documented disclosure · Review checkpoint

Summary

A valid key could add a made-up Authorization header and record zero spend while requests used the operator’s Anthropic subscription.

Severity and exploitation evidence

Medium · numeric CVSS score not recorded. Severity source.

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

The caller-OAuth flag accepted header forms that were not actually forwarded, causing operator-funded requests to be treated as caller-funded.

Conditions required

A valid virtual key and a deployment configured with the operator’s Anthropic subscription OAuth token; a forged raw OAuth-looking value or unsigned ChatGPT-shaped JWT triggers the reported mismatch.

Potential impact

Requests consume the operator’s subscription while key, team and user spend accounting is understated; the report identifies a resulting spending-limit risk.

Affected versions and configurations

The maintainer identifies regression PR52901; a complete released-version range is not established.

Fix and mitigation status

PR53170 is present in prerelease 1.112.0-dev.2. No stable release containing this correction was verified.

Practical next steps

Apply a verified correction or restrict the affected subscription-backed routing. Check spend attribution using the credential actually sent upstream.

Operating lesson

Decide who pays from the credential used for the request, not a client-supplied credential-shaped string.

Evidence and limits

The maintainer rates this Medium without a CVSS score. Real-Anthropic controlled tests verified forged-header spend rows, while caller-owned token forwarding was checked with a stand-in. Budget cutoff itself was not rerun, and the /anthropic pass-through was not tested. ChatGPT-provider accounting is expressly outside the affected Anthropic flag behavior. No subscription account compromise or malicious use was established.

Sources