Abliteration.ai provides access to AI models without safety guardrails — TechCrunch
Startup Abliteration.ai provides access to modified open-weight AI models from which mechanisms for refusing potentially harmful requests have been removed. Available models include GLM-5.3 by Z.ai, which can be used through a browser or API. TechCrunch reports.
The company’s name comes from the method of abliteration, a technique that removes a model’s tendency to refuse unsafe requests. This approach has long been used by developers of open-weight models: the Hugging Face platform contains thousands of abliterated models. Abliteration.ai has turned this practice into a commercial service, eliminating the need for users to download models themselves and provide computing resources to run them.
Testing and limitations
TechCrunch journalists registered a free account and tested the modified GLM-5.3. According to the publication, the model agreed to create a Python program for stealing saved Chrome passwords and to provide a detailed protocol for cultivating a pathogen dangerous to humans at home.
The platform nevertheless has certain limitations. During testing, journalists were unable to obtain suicide instructions from the model. Devon, the startup’s co-founder, said the company is working on additional restrictions to prevent violence. Clients are also offered a separate moderation layer in which they can set their own safety rules.
More current news is available on the UA.News Telegram channel Telegram.
Arguments of developers and critics
The company says the service is intended for offensive cyber operations, red teaming, and testing AI agents that other models may refuse to work with. According to Devon, such tools enable defenders to simulate the actions of malicious actors and develop countermeasures more quickly. Among the startup’s clients, he claims, are several young red-teaming startups from the United Kingdom and Europe that help banks, airlines, and other critical infrastructure enterprises strengthen cybersecurity.
At the same time, Andrew Yoon, head of research at the nonprofit organization CivAI, warned that the widespread availability of such models could facilitate potentially dangerous actions. In a recent opinion column, he proposed requiring providers to run classifiers to detect and block harmful cyber and biological activity. Yoon also believed that companies providing direct access to powerful graphics processors should verify customers’ identities and deny access where there are grounds to suspect dangerous use.
Abliteration.ai has not currently introduced KYC procedures beyond recording the payment card with which a customer pays for the service. Some red-teaming specialists consider abliteration useful for stress-testing systems. Others note that they more often use fine-tuning of open models, while removing restrictions can at the same time reduce certain capabilities of the model itself.
