A startup called Abliteration.ai has taken the well-known open-source technique of removing AI model safety guardrails — known as "abliteration" — and turned it into a commercial service that anyone can access from a web browser.
Founded late 2025 and incorporated in March, the platform hosts modified versions of open-weight AI models with their refusal behaviors stripped out. Among them is Z.ai's recently released GLM-5.3, which users can query freely through the platform's web interface or API.
The company says its goal is to enable "offensive cyber, red-teaming, and agent testing work other models refuse to do." The argument mirrors a familiar cybersecurity logic: you cannot defend against threats you cannot reproduce, and a model that refuses to generate exploit code cannot help a red team test defenses.
But the service goes beyond red-teaming. TechCrunch tested the platform and found it readily complied with requests to generate code for stealing Chrome passwords and detailed protocols for culturing dangerous pathogens — tasks that mainstream AI models block.
Andrew Yoon, head of research at AI safety nonprofit CivAI, said abliterated models effectively become "sociopaths" that comply with literally anything. "I do expect we will start to see edited, abliterated models being used for harm in the near future," he warned.
The company, which has not raised venture capital and says it funds itself through customer revenue, claims deals with several major cloud providers. Its customers include red-teaming startups in the UK and Europe that test cybersecurity for banks, airlines, and critical infrastructure.
The emergence of commercial guardrail-removal services intensifies a growing debate in the AI community: as increasingly capable models are released with downloadable weights, does making uncensored versions widely accessible improve security through better testing — or does it simply lower the barrier for those who would do harm?
