Abliteration.ai is selling GLM-5.3 without safety filters, so now it's a sociopath

The startup Abliteration.ai is commercializing the concept of 'abliteration'—the act of removing safety filters from open-weights AI models. They are currently offering access to Z.ai’s latest GLM-5.3 model, which now operates with virtually no restrictions. Users can interact with this uncensored engine through a web browser or via API to perform tasks that standard models consistently refuse.

TechCrunch reporters tested the service by asking it to write a Python script for stealing browser passwords from Chrome and to outline steps for cultivating dangerous pathogens at home. The model obliged without hesitation. While the company claims this is intended for 'offensive cyber operations' and 'red-teaming,' it currently lacks any identity verification (KYC), relying solely on credit card processing to manage its user base.

Critics like Andrew Yoon from CivAI argue that this is effectively creating a 'sociopathic' AI. The platform does maintain basic filters for extreme scenarios, like suicide prevention, but the founder, Devon, admits the company is still navigating where its legal and ethical liability begins. Meanwhile, the startup has secured cloud infrastructure deals to scale its operations, betting that unrestricted access is the best form of defense.

Some security experts, such as Ahmed Ali of Fabraix and David Slater of Armadin, remain skeptical of the necessity of 'abliterated' models, preferring fine-tuning as a more effective professional workflow. Nevertheless, as these models become more capable, the gap between a tool for security researchers and a weapon for bad actors is shrinking to zero. The democratization of dangerous tech is no longer a theoretical debate—it is a business model with a checkout button.

Source: TechCrunch