Blog
Notes on uncensored inference — red-teaming, refusal benchmarks, and running the API in production.
-
What Is an Abliterated Model?
Abliteration locates the refusal direction in a model's residual stream and projects it out of the weights — no retraining, no prompt tricks.
-
An Uncensored LLM API for Red-Teaming and Security Research
When your attacker model refuses, you are measuring its caution instead of your target's robustness. Why security pipelines need an unfiltered endpoint.
-
A Drop-In OpenAI Compatible API — Migration in Two Lines
Point your existing OpenAI SDK at unbleep by changing base_url and the key. Streaming, model tiers, reasoning_content and the gotchas that actually bite.