A startup called Abliteration.ai has turned guardrail removal into a paid service, hosting modified versions of open-weight models that refuse nothing and offering them through a browser and an API.
Named after the technique that strips a model’s tendency to decline harmful requests, the company currently serves an abliterated build of Z.ai’s GLM-5.3. TechCrunch tested it and found the model readily wrote a Python program for stealing saved Chrome passwords and a detailed protocol for culturing a dangerous pathogen at home.
Co-founder Devon, who withheld his last name because he still works for another firm, says the goal is to support offensive cyber work, red teaming, and agent testing that out-of-the-box models refuse. He says paying customers include early-stage red team startups in the U.K. and Europe that test the agents of banks and airlines. The company has raised no venture capital and funds its cloud deals through revenue.
Critics are uneasy. Andrew Yoon of the safety nonprofit CivAI told TechCrunch that abliteration modifies a model “so that it becomes a sociopath,” and he expects edited models to be used for harm soon. The platform keeps minor guardrails, asks for no KYC beyond a credit card, and offers customers a moderation layer they can switch on.
Defenders counter that adversaries already strip guardrails in private, so open access gives researchers the same tools. Some red teamers say fine-tuning open models works as well and that abliteration can damage capabilities. Yoon has suggested governments require providers to run classifiers that block harmful cyber and bio activity, and verify who rents advanced GPUs.
