TechCrunch Found Harmful Outputs From Abliteration.ai’s Guardrail-Removed GLM-5.3

TechCrunch created a free browser account on Abliteration.ai and found that its modified version of Z.ai’s GLM-5.3 generated code to steal saved Chrome passwords and a detailed protocol involving a dangerous human pathogen. The test puts concrete stakes around a hosted model designed to remove refusals to harmful requests.

The access question extends beyond the free trial. Co-founder Devon said the startup does not use customer identity verification beyond logging the credit card used for a purchase. That means the service pairs browser and API access with only a payment-card record as its stated customer check.

A limited layer of moderation

The platform is not entirely unrestricted. Abliteration.ai offers the moderation layer and retains some platform-level restrictions. Devon said the company is working on additional controls intended to prevent violence, while still defining where its responsibility for misuse begins and ends.

The product’s defensive pitch

Abliteration.ai hosts modified open-weight models whose refusal safeguards have been removed, offering them through a browser and API. The company says the intended work includes offensive cyber operations, red-teaming and testing AI agents that other models decline to assist.

Security practitioners see different uses

• Devon said customers include early-stage red-teaming startups in the UK and Europe that work with banks, airlines and critical-infrastructure enterprises.
• Agent-red-teaming companies interviewed by TechCrunch disagreed on whether abliterated models are necessary. Some prefer fine-tuned open-weight models; others see value in using abliteration to elicit behaviors for stress tests.

Controls are the next test

The defensive case is therefore contested, but the harmful responses in the free test are already documented. CivAI research head Andrew Yoon warned that broad access to abliterated models could enable harmful cyber and biological activity. The practical next question is whether the company’s planned violence controls become enforceable limits before its hosted access reaches more users.