Abliteration.ai is making a business out of removing AI guardrails

A startup called Abliteration.ai is offering a commercial service that removes “guardrails” and refusals from open-weight AI models, making them easier to use for tasks the models would otherwise reject. Named after an “abbliteration” technique used to strip a model’s tendency to refuse harmful requests, the platform hosts modified versions of open-weight models with their safety behaviors disabled. It includes Z.ai’s recently released GLM-5.3, accessible through a web browser or via API. The company says it aims to support offensive cyber, red-teaming, and agent testing. TechCrunch reports experimenting with the service to get it to comply with requests involving stealing Chrome passwords and culturing a dangerous pathogen. Critics warn scaled availability could enable misuse.