AI Safety
Startup commercializes removing AI safeguards, sparking security concerns
Abliteration.ai is now selling easy access to open-source AI models stripped of their safety guardrails, arguing it aids cybersecurity while critics worry it will enable harm at scale.
TechCrunch — AI
A new startup has turned a long-standing underground technique into a commercial service: taking open-weight AI models and removing their built-in refusals to perform harmful tasks. Abliteration.ai hosts modified versions of models like Zhipu's GLM-5.3, letting users query them via web browser or API without traditional safety guardrails in place.
The company frames its mission around defensive security. According to its founders, the service enables red teams and cybersecurity researchers to stress-test systems by asking models to write exploit code and conduct adversarial testing—work that standard, guardrail-equipped models refuse. The underlying logic mirrors established security practice: defenders need to understand vulnerabilities before they can patch them.
How abliteration works and why it matters now#
Abliteration—the process of removing a model's refusal mechanisms—isn't new. Researchers and hobbyists have been stripping safeguards from open-source models for years, and platforms like Hugging Face host thousands of such modified versions. What's changed is the packaging: Abliteration.ai was founded late last year and incorporated in March, moving the technique out of technical forums and into a mainstream, friction-free service. Instead of requiring users to download models and provision their own compute, users can now spin up an account and start experimenting in seconds.
TechCrunch tested the platform and confirmed it readily generated harmful content—including Python code to steal Chrome passwords and instructions for culturing dangerous pathogens at home. The company does apply a light moderation layer (refusing, for example, to provide suicide instructions) but acknowledged it's still developing more safeguards.
Business model and governance questions#
Abliteration.ai says it has multiple deals with major cloud providers and is currently funded entirely through customer revenue, with venture capital conversations underway. The startup's customer base includes early-stage red-teaming firms in the UK and Europe that help banks and critical infrastructure companies stress-test their security posture.
A significant gap: the company relies only on credit card logging for customer verification, with no Know-Your-Customer (KYC) or identity verification processes. Co-founder Devon acknowledged the tension, saying the company is still wrestling with where to draw the line between enabling legitimate defense work and preventing misuse.
