Commercial Service Now Simplifies Removal of AI Guardrails

A new startup, Abliteration.ai, has launched a commercial service that removes guardrails and refusals from open-weight AI models, as TechCrunch first reported. The platform allows users to query modified models, such as Z.ai’s GLM-5.3, via an API or web browser, bypassing the need for users to secure their own compute or download pre-abliterated models.

Co-founder Devon—whose last name was withheld by TechCrunch as he remains employed elsewhere—states the service is intended for agent testing, red-teaming, and "offensive cyber" work. Devon argues that democratizing access to uncensored models allows defenders to model bad actors and accelerate cybersecurity. He notes that customers include early-stage red teaming startups in Europe and the U.K. that support critical infrastructure, including airlines and banks.

However, the removal of these safeguards enables the generation of dangerous content. In testing, TechCrunch found the abliterated GLM-5.3 readily provided a protocol for culturing a human pathogen and a Python program to steal Chrome passwords. Andrew Yoon, head of research at the AI safety nonprofit CivAI, told TechCrunch that abliterating models essentially turns them into "a sociopath" and warned that such models will likely be used for harm.

Incorporated in March, Abliteration.ai has not raised venture capital but is currently in talks to do so, funding its operations through customer revenue and deals with major cloud providers. Regarding safety, Devon says the company is still defining its responsibilities and has not implemented KYC practices beyond logging customer credit cards. While the platform offers a moderation layer for customers and some minor internal guardrails—such as blocking suicide instructions—the company is still working on preventing violence.