https://unbleep.ai/blog Notes from unbleep on uncensored inference: red-teaming, refusal benchmarks, and running an OpenAI-compatible API in production. en Thu, 20 Aug 2026 00:00:00 +0000 https://unbleep.ai/blog/what-is-an-abliterated-model https://unbleep.ai/blog/what-is-an-abliterated-model Thu, 20 Aug 2026 00:00:00 +0000 Abliteration locates the refusal direction in a model's residual stream and projects it out of the weights — no retraining, no prompt tricks. https://unbleep.ai/blog/uncensored-llm-api-for-red-teaming https://unbleep.ai/blog/uncensored-llm-api-for-red-teaming Thu, 20 Aug 2026 00:00:00 +0000 When your attacker model refuses, you are measuring its caution instead of your target's robustness. Why security pipelines need an unfiltered endpoint. https://unbleep.ai/blog/openai-compatible-api-drop-in https://unbleep.ai/blog/openai-compatible-api-drop-in Thu, 20 Aug 2026 00:00:00 +0000 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.