Abliteration.ai has launched a commercial service that removes trained refusal mechanisms from open-weight AI models. The company's latest offering, abliterated-model-large-v2, is a modified version of Z.AI's GLM-5.3, designed to reduce the frequency of sensitive request refusals.
The process, known as abliteration, involves identifying and suppressing internal activation patterns that trigger refusals by tweaking model weights. Abliteration.ai claims that coding, cyber, and agentic capabilities remain largely intact. For the modified GLM-5.3, the company reports 84.5% on CyberGym, 41.8% on Terminal-Bench 4.0, and 105 solved ExploitGym tasks in two hours.
The service is priced at $5 per million input or output tokens. Unlike some developers who publish modified weights on Hugging Face, Abliteration.ai hosts the model and manages the operations, providing API access without requiring users to download the model or maintain GPU infrastructure.
The company markets the model for offensive cybersecurity, AI red teaming, and malware analysis, including simulating phishing attacks and reproducing vulnerabilities. However, TechCrunch reported that the model produced code for extracting Chrome passwords and a guide for creating a dangerous pathogen, though it still blocks self-harm and child sexual abuse material.
Abliteration.ai does not store prompts and responses, which protects confidential source code for security teams, but makes it harder to investigate abuse. The company does not require identity verification for standard access. Enterprise customers can use an optional policy gateway to define request rules, and the company also sells synthetic training and evaluation data.
The company is registered for US government procurement on SAM.gov. Other providers like Audn.AI and Silk Compute also host unrestricted models for security purposes. A preliminary study suggests that while refusals can be reduced, abliteration may cause behavioral changes even in tasks where the base model did not refuse.
