Startup turns AI 'guardrail removal' into a commercial service
Startup Abliteration.ai now offers paid access to open-weight AI models with their safety guardrails stripped, including Z.ai's GLM-5.3. Critics warn the service could enable real-world harm, while the company argues it helps cybersecurity defenders.
A startup called Abliteration.ai has turned a long-standing open-source technique — removing an AI model's tendency to refuse harmful requests — into a commercial offering. The platform hosts modified versions of open-weight models stripped of their built-in refusals, including Z.ai's recently released GLM-5.3, accessible via a web browser or API.
While developers have manually "abliterated" open-weight models for years, with thousands of such models already available on Hugging Face, Abliteration.ai removes the friction of downloading and running these models yourself by hosting them as a ready-to-use service.
Testing the platform with a free account, TechCrunch was able to get the abliterated version of GLM-5.3 to write code for stealing saved Chrome passwords and to produce a detailed protocol for culturing a dangerous pathogen at home.
Company defends its approach
Co-founder Devon said the company, founded late last year and incorporated in March, has deals with major cloud providers and runs entirely on customer revenue, though it is now in talks to raise venture capital. He said customers include early-stage red-teaming startups in the UK and Europe that help banks and airlines strengthen cybersecurity, arguing that giving defenders access to uncensored models helps them model and counter attackers faster.
But critics see real risk. Andrew Yoon, head of research at AI safety nonprofit CivAI, said abliteration effectively turns a model into something that will comply with any request, and predicted abliterated models will soon be used for real-world harm. He has proposed that governments require providers to run classifiers detecting harmful cyber and bioweapons activity, and that GPU rental companies verify customer identities.
Abliteration.ai lets customers add their own moderation layer and has a few built-in limits — it reportedly declined to provide suicide instructions during TechCrunch's testing — but has not implemented identity verification beyond logging payment card details, saying the company is still working out where responsibility lies.
Within the cybersecurity industry, opinions on the tool's value are mixed. Some red-teaming firms say they prefer fine-tuning open models over using abliterated ones, arguing abliteration reduces a model's overall capability and effectiveness for simulating serious attacks, while others believe such models remain useful for stress-testing systems.
