Is text from audio files via "Upload as Text" & mic external prompt leaking across all users for anybody else? 0.8.0

[gmerritt]

When I attach an audio file via "Upload as Text", using whisper Azure via litellm, the text comes back fine. But then any subsequent upload of any different audio file of the same extension (must match, .m4a or .mp3 or whatever), the first file's text is returned... and this is independent of user and browser!!

And same goes for External STT -- User A in Chrome uses STT mic, "Hello, what do you do?" and it works; user B accesses via Safari, speaks "What time is it in Paris?" and gets the first user's "Hello, what do you do?" text. All users get that first-in-line mic prompt's text.

It all resets after some fraction of an hour, and then repeats with the first audio submission of the same type.

Anyone else experience this?

Note that STT via mic has been working fine on 0.7.9 for us.

I'll file a bug report next week with decent documentation if there are no other leads before then. :)

Thanks!!

[gmerritt]

Ohhh whoops! It is happening in our pre-0.8.0 for us too smh....this is why I haven't filed an issue yet, I need good documentation...I'll also check our litellm.

[Odrec]

We have experienced some weird leakage but not from audio files. gpt-5 thought it was Gemma3 because we have a custom endpoint with Gemma3 with a custom system prompt. I'm not sure why this happens.

[gmerritt]

I guess my use of "leakage" was as in "my private data is leaking to you in your account"

[gmerritt]

Ok so...I am now suspicious of litellm, sigh...I will report in this thread what I ultimately confirm.