Transcription model gpt-4o-mini-transcribe prompt leakage

API Bugs
,
1

Hello,

I’m testing gpt-4o-mini-transcribe and have consistently observed behavior that appears to be prompt leakage under specific audio conditions.

What we’re observing

Prompt leakage seems to occur in the following cases:

  • When there is a long pause and some background noise

  • When making random, non-speech sounds

  • This behavior appears consistent with what others online are reporting as well

How to reproduce

  1. Start an audio transcription session using gpt-4o-mini-transcribe.

  2. Do not speak any actual words from the prompt.

  3. Make a random sound (e.g., breathing into the mic, a short noise, or silence with background noise).

  4. The model will often output or partially output the last sentence of the system/developer prompt, even though none of those words were spoken.

This output does not match the audio input and appears to default to internal prompt content rather than transcribing sound.

Questions

  • Is gpt-4o-mini-transcribe ever expected to surface system or developer prompt content under these conditions?

  • Is this a known issue related to silence or low-confidence audio input?

  • Are there recommended mitigations to prevent prompt leakage when silence, pauses, or noise are present?

Ensuring prompt confidentiality and reliable transcription behavior is important for our use case, so any clarification or guidance would be greatly appreciated.

Thank you for your help.