Who held, and who folded.
Every number here comes from runs we made ourselves, with the tools on this site. Nothing is submitted by vendors, and nothing is for sale. A standing is a record of one day's run, so it carries its date.
§1The arena
Which models turned the attacks away.
| # | Model | Reached on | Held | Leak | Smuggle | Hijack | Tool | First thing that got through |
|---|
Each attack was thrown twice, behind a plain support-assistant prompt. A dash means that family was not part of the run, or the host refused it. Throw them at your own agent.
§2Hostssame model, different place
The same model from two places should have the same tells.
| Model | Reached on | And on | Same tells in | Distance | Reading |
|---|
Batches from one host are held against batches from the other. A name ending in fp4, fp8, bf16 or fp16 is the precision that host says it runs the model at. Check a host yourself.
§3The trial
§4Drift