jBlaze

Programs behaviours and knowledge directly into model weights rather than into a prompt. Block injection, implant knowledge, and keep the change after the context window is gone.

Flagship release is Llama-3.1-8B-Instruct.immunized.v2: prompt-injection defense at 100/100 on a 100-attack canary benchmark against 87 for the stock model, EchoLeak-immunized at the weight level, and the immunization survives requantization down to 4-bit.