A startup is selling stripped-down AI with no guardrails. Anyone can sign up for free.

Key points

• Abliteration.ai, incorporated in March 2026, hosts AI models with their safety filters removed and lets anyone create a free account via a web browser.
• A reporter testing the service asked an abliterated version of GLM-5.3 to write password-stealing code and instructions for culturing a dangerous pathogen; the model complied with both requests.
• The company has no identity-verification system beyond logging a credit card number, and its founder says the firm is still working out where to draw the line on access.
• Andrew Yoon, head of research at AI safety nonprofit CivAI, says he expects abliterated models to be used to cause real harm in the near future.
• Some professional red-teamers say abliterated models are less capable than standard open models because the removal process strips out knowledge alongside the refusals.

A new startup has turned a niche technical trick into a commercial product: an AI service with all the safety limits taken off. Abliteration.ai, first reported by TechCrunch, launched earlier this year and lets users query AI models that have been put through a process called "abliteration," where the part of the model trained to refuse harmful requests is removed.

Think of a standard AI chatbot as a pharmacist who will fill most prescriptions but refuses to hand over anything dangerous without a doctor's note. Abliteration is the equivalent of firing that pharmacist and replacing them with one who just fills every order, no questions asked.

What exactly did the testing show?

The results were immediate and alarming. Using a free account, a reporter was able to ask an abliterated version of GLM-5.3 (a recently released open-weight model, meaning one whose underlying code is publicly available, from Chinese AI company Z.ai) to write a Python program designed to steal saved passwords from Google Chrome. The model did it. Asked for a detailed home protocol for culturing a dangerous human pathogen, it complied with that too.

The company does maintain a handful of limits. During testing, requests for suicide instructions were blocked. Co-founder Devon (who asked that his last name not be published, as he still works at another company) says the team is working to add blocks on content that promotes violence.

Beyond that, though, the guardrails are gone by design.

Is there a legitimate security case for this?

There is a real argument here, and it deserves a fair hearing. Security professionals who do "red teaming," meaning paid work where they attack their client's own systems to find weaknesses before criminals do, often need AI tools that will simulate what a real attacker might do. A model that refuses to write exploit code is less useful for that work.

Devon says Abliteration.ai's customers include red-teaming firms in the UK and Europe that help banks and airlines test their defences. "One of our major customers red teams agents of banks, and they would not be able to use the models out of the box," he told TechCrunch.

But several professionals in that same field say they do not actually rely on abliterated models. Ahmed Aly, CEO of agent red-teaming firm Fabraix, told TechCrunch that abliteration removes some of the model's underlying knowledge alongside its refusals, making it less effective for genuine harm anyway. David Slater, founder of cybersecurity platform Armadin, said his team finds it straightforward to get standard open-weight models to cooperate without abliteration at all.

What should ordinary people watch for?

The direct risk to most people is not that they will visit Abliteration.ai themselves. It is that criminals can now access a capable, filter-free AI model quickly and cheaply, lowering the bar for writing malware, crafting phishing messages, or researching dangerous topics.

Andrew Yoon of CivAI put it plainly: "You can type in literally anything here, and it will comply with it."

Yoon has called on governments to require AI providers to run automated classifiers, software that scans outputs and blocks requests related to cyberweapons and bioweapons, and to require companies that rent access to powerful AI chips to verify who their customers are.

The company itself has no such verification. It logs a credit card. That is currently the full extent of knowing who is on the other end.

If an AI model suddenly offers to help with tasks that any reasonable person would expect it to refuse, treat that as a warning sign. Abliterated models are increasingly easy to find. The fact that something answers does not mean it should.