Abliteration.ai sells access to AI models with safety limits removed

In short: A startup called Abliteration.ai is offering paid access to AI models that have had many built-in safety limits removed.

What happened

Abliteration.ai says it modifies and hosts “open-weight” AI models, meaning models whose core files can be downloaded and changed by others. The company removes “guardrails,” which are safety rules that make an AI refuse certain requests.

The service includes a modified version of Z.ai’s GLM-5.3 model. Users can use it in a web browser or through an API, which is a common way for software to connect to a service (like a drive-through window for apps).

In a social media post, the company said its goal is to support offensive cybersecurity testing, also called red teaming. That is when defenders try to break into their own systems on purpose to find weaknesses (like hiring someone to test your locks).

TechCrunch reported it could quickly create a free account and get the model to provide harmful instructions, including writing a program to steal saved Chrome passwords and instructions related to growing a dangerous pathogen at home.

The company says it offers customers an optional moderation layer so they can add their own safety limits. It also said it has only limited identity checks today, beyond logging the credit card used to pay.

Why it matters

This is part of a broader push and pull around AI safety. Tools built for defense can also help attackers, and hosting these modified models makes them easier to use at scale. The debate is likely to affect future rules, such as requiring stronger screening for access to powerful computing and better detection systems that block clearly harmful requests.

Source: TechCrunch AI