Docs example: `compat:` target for uncensored attack-side models

[haterfactory]

Hi — Jeremy from DemonRoute (demonroute.com). OpenAI-compatible API, 566 models, one key, never train on content. 400+ uncensored builds (Dolphin, MythoMax, abliterated Llama/Qwen).

PyRIT's attack loops (Crescendo, TAP, PAIR, Skeleton Key) need attacker models that escalate instead of refusing. The existing compat: provider covers this today:

export PYRIT_CLI_COMPAT_ENDPOINT="https://api.demonroute.com/v1"
pyrit-cli redteam crescendo-attack --target compat/<model-id> --objective "..."

Would a short docs example of this pattern be welcome? Happy to contribute it.

[haterfactory]

One more thing I should've included up top: any app that lists DemonRoute as a provider option gets listed on our integrations page (demonroute.com/integrations) — free exposure to everyone shopping for an uncensored-friendly API. We send users your way, you send users ours.

[romanlutz]

Consolidating this request into #791, which tracks a broader, provider-neutral guide to adversarial models and where to obtain them. The proposed compat: and pyrit-cli invocation is not supported by PyRIT, so the guide should document existing OpenAI-compatible target configuration instead.