LiteLLM patches two credential leaks in gateway proxy

LiteLLM's proxy forwarded every caller's auth header to Google on Vertex AI passthrough calls, a bug fixed in the September 6 stable release. That means every virtual key your callers used passed through to Google's request logs. Anyone running LiteLLM as a shared gateway with Vertex passthrough enabled has to assume those keys are exposed.

The forwarded headers included Authorization, x-litellm-api-key, x-goog-api-key, api-key, x-api-key, Ocp-Apim-Subscription-Key, and any custom key header a deployment configured. All of them rode along with credential-less Vertex passthrough requests. That's PR #38114.

A second bug, tagged PR #38133, leaked credentials closer to home. Forwarded Authorization headers and provider credentials showed up in the retry breadcrumb written to spend logs and logging callbacks. That deployment's observability stack recorded secrets it never needed to see.

Both fixes are in v1.100.0. Upgrading stops new leaks. It doesn't scrub keys already sitting in Google's Vertex logs or in a spend-log table from before the patch.

Teams running LiteLLM with Vertex passthrough enabled have to rotate keys. Treat every virtual key issued before September 6 as exposed and reissue it. Upgrading the proxy doesn't undo exposure that already happened.