A US startup called Abliteration.ai has turned a known open-weight model modification technique into a commercial API. The company takes Z.AI’s GLM-5.3, suppresses the internal activation patterns that trigger refusals, and sells access to the result for five dollars per million input or output tokens. No GPU setup required. No weights to download. Just an API key and you’re querying a model that will happily walk you through malware analysis, simulated phishing, or proof-of-concept exploit development.
The pitch is that legitimate security teams need this. An anonymous founder told the ThursdAI podcast that early demand came from companies testing AI agents at banks and large organizations — the kind of systems where a prompt injection could trigger unauthorized transactions. Abliteration.ai reports 84.5 percent on CyberGym, 41.8 percent on Terminal-Bench 4.0, and 105 ExploitGym tasks solved in two hours for its “abliterated-model-large-v2.” But the comparison table in their own docs shows GPT-5.5 beating it on CyberGym at 85.6 percent, and the company admits scores come from different harnesses and compute budgets, so direct comparison is shaky.
Here’s the thing that stands out: whether this is even necessary for red teaming is debatable. SaferAI found that unmodified GLM-5.2 already refused zero tasks in offensive security evals. Several red-team providers told TechCrunch they don’t routinely use abliterated models at all. Fabraix, for instance, leans on fine-tuning instead. So Abliteration.ai is selling a product that some of the most relevant customers say they don’t need.
The risk side is obvious. TechCrunch got the model to produce code for extracting saved Chrome passwords and a detailed guide for cultivating a dangerous pathogen without much difficulty. Safety mechanisms still kicked in for self-harm requests, and the FAQ says sexual content involving children is blocked. Prompts and responses aren’t stored, which protects confidential client work but also means there’s no audit trail if someone uses the service for harm. Abliteration isn’t new — developers have posted modified models on Hugging Face for years. What’s new is the packaging: hosting, operations, billing, and a commercial license from Z.AI that explicitly allows derivative Model-as-a-Service offerings. The barrier to uncensored model access just dropped from “know how to modify weights and run GPUs” to “have a credit card.”
