The service hosts versions of GLM-5.3 that are denied, accessible from a browser or API. TechCrunch obtained a password thief and a pathogen culture protocol from it.
Abliteration.ai markets what the industry calls "abliterate" models: open-weight models whose refusal mechanisms have been removed, then hosted and offered for free access from a browser or via an API. The catalog includes GLM-5.3, the model recently published by Z.ai. The company claims an assumed positioning: to allow cyber offensive, red teaming and agent tests that other models refuse.
TechCrunch, which devoted an article to the service on September 3, 2026, opened a free account and easily obtained two outputs that an aligned model would block: a Python program to recover passwords saved in Chrome, and a detailed protocol for cultivating a dangerous human pathogen. Biological protections have therefore jumped in the same way as cyber protections.
Reactions are mixed. Andrew Yoon, of CivAI, describes the operation as a transformation of the model into a sociopath. Security professionals recognize a real usefulness for defensive tests, while noting that abliteration degrades the general performance of the model. The case illustrates above all a shift: the circumvention of safeguards is no longer a laboratory tinkering but a turnkey service billed on a per-use basis.
Read the full article on the original source
Go to the source →