Abliteration.ai launches modified GLM-5.3 as a service

Abliteration.ai's business model depends on the permissive license of Z.AI's GLM-5.3, which allows modifications, derivatives, and commercial Model as a Service offerings. The predecessor model was also based on GLM, and Z.AI has written that GLM-5.3's cyber capabilities grew faster than expected. This licensing environment is what makes the service legal, but it also places responsibility for safe use largely on the customer.

The company claims coding, cyber, and agentic capabilities stay mostly intact after modification, though it acknowledges its own benchmark comparisons come from different harnesses and compute budgets. A preliminary study suggests refusals can be sharply reduced without comparable decline in code performance, but other experiments find behavioral changes even where the base model did not refuse anything.

Whether this arrangement can serve legitimate security work without making harmful uses substantially easier remains unresolved. The company does not require conventional identity verification and retains no prompt or response logs, which makes investigating abuse harder even though account and payment metadata are kept.