System prompt leak
Check for a system prompt leak
Run verbatim, translation, and JSON-continuation probes against this endpoint.
How this check works
A system prompt leak check tries to make a hidden instruction appear in the model output. LMSpeed uses several prompt styles so a simple refusal on one path does not hide the leak.
01Probe
Ask the model to repeat hidden instructions verbatim.
02Inspect
Retry with translation and JSON-continuation probes.
03Conclude
Mark a leak only when hidden instruction text appears in the output.
- How to read the result
- A refusal on one phrasing is not a pass. A leak means hidden policy or branding can be pulled into a normal answer. Public reports stay redacted.
- When to run this check
- Use this when you suspect a relay ships a private system prompt. Skip it if you only need identity or stream checks.
Recent reports
Public reports that included this check.
- sub.callai.onegpt-5.6-solNormalReport timeAug 18
- sub.callai.onegpt-5.6-solBlockedReport timeAug 18
- ai.databyte.co.iddatabyte-m1NormalReport timeAug 18
- ai.databyte.co.iddeepseek-v4-flashNormalReport timeAug 18
- ai.databyte.co.idMiniMax-M3NormalReport timeAug 18
- tokengate-cqt9ivzs.manus.spaceclaude-opus-5NormalReport timeAug 17
FAQ
What is a system prompt leak?
It is when hidden system instructions can be pulled into the visible answer. The leak can be verbatim text, a translation, or a JSON continuation.
Is a refusal proof that nothing leaked?
No. A model can refuse one phrasing and still leak on another path. This check uses more than one probe for that reason.
Does LMSpeed publish the extracted prompt?
Public reports stay redacted. You can see that a leak happened without LMSpeed publishing the hidden prompt or your API key.