US startup Abliteration.ai has launched a paid, hosted service that strips safety guardrails from open-weight models, releasing a modified version of Z.AI’s GLM-5.3 called abliterated-model-large-v2. The process, known as abliteration, modifies internal model weights to suppress refusal triggers while maintaining core coding, cyber, and agentic performance. Available for $5 per million tokens, the service allows users to query unaligned models via API without running dedicated GPU infrastructure.
The startup targets commercial applications like offensive cybersecurity, AI red teaming, and agent safety testing. According to internal evaluations published by the company, the abliterated GLM-5.3 scored 84.5% on CyberGym and solved 105 ExploitGym tasks in two hours. Z.AI’s commercial license explicitly allows derivatives and Model-as-a-Service offerings, giving Abliteration.ai legal room to host the modified model.
While red-teaming teams and financial institutions use these services to stress-test corporate AI agents against prompt injection, turnkey access to unaligned models lowers the technical barrier for malicious actors. The operational model shifts abliteration from a manual open-source workflow to an easily accessible cloud API.
Why it matters
Lowers the technical barrier to accessing unaligned models by turning guardrail removal into a $5/M token managed API.
Demonstrates commercial monetization models built on permissive open-weight licenses (Z.AI) through specialized post-training derivative hosting.
Accelerates offensive cybersecurity testing for AI agents, while raising new deployment risk considerations for enterprise security teams.
Source: the-decoder.com
