Abliteration.ai launches offensive cybersecurity large model based on GLM-5.3

PANews reported on September 1 that the startup Abliteration.ai announced the launch of a new model, "abliterated-model-large-v2," based on GLM-5.3, which ranked third on Terminal-Bench 4.0. The company claims its network attack and defense exploitation capabilities are approximately 2 times better than those of the 5.2 version. The official statement said that by removing the "refusal" direction in model activations at the weight level, while retaining coding, cybersecurity, and agent capabilities, the model can be used in offensive cybersecurity, AI red team assessments, agent testing, and trust and safety scenarios to fully execute authorized attack chains. The model is hosted in the United States, uses FP8 precision, supports a 1 million-token context window, and promises not to retain input and output prompts. It is currently available to the public via API.