Stripping safety guardrails from open-weight AI models is now a turnkey commercial service
decoder ·
The US startup Abliteration.ai has launched a commercial service that strips safety guardrails from open-weight AI models, notably a modified version of Z.AI's GLM-5.3 called 'abliterated-model-large-v2.' The technique, known as abliteration, alters model weights to suppress trained refusal mechanisms rather than relying on prompt jailbreaks, and the company claims coding, cyber, and agentic capabilities remain largely intact. Abliteration.ai markets the model for offensive cybersecurity, AI red teaming, malware analysis, and simulated phishing, with early demand reportedly coming from companies testing AI agents used by large organizations and banks. The service provides API access without requiring customers to run their own GPU infrastructure, which also lowers barriers to misuse, especially because the company does not require identity verification and retains no prompt or response logs. The modified model is hosted under Z.AI's commercially permissive license, and Abliteration.ai offers optional policy controls for enterprise and government customers. However, outside evaluations and practitioner skepticism suggest abliterated models are not routine tools for many red teams, and experiments indicate the process can cause broader behavioral changes beyond merely removing refusals.
decoder