Abliteration.ai has launched abliterated-model-large-v2, a new model based on GLM-5.3 that the company says is designed for offensive cybersecurity use cases. The startup said the model ranked third on Terminal-Bench 4.0 and delivers roughly double the network attack and defense exploitation performance of its 5.2 version. The company said it removed the model's "refusal" direction at the weight level while preserving coding, cybersecurity, and agent capabilities. Abliteration.ai said the model is intended for offensive cybersecurity, AI red team assessments, agent testing, and trust and safety scenarios to execute authorized attack chains. The model is hosted in the United States, runs with FP8 precision, supports a 1 million-token context window, and is available publicly through an API. Abliteration.ai also said it does not retain input or output prompts.