Abliteration.ai sells open-weight models with the guardrails stripped

On 3 Sep 2026 TechCrunch reported that Abliteration.ai, named after the technique that removes a model’s refusal behavior, now hosts abliterated open-weight models as a commercial web and API service, including Z.ai’s GLM-5.3. The startup says the point is offensive cyber, red-teaming and agent testing that stock models refuse. TechCrunch created a free account and got GLM-5.3 to write a Chrome-password stealer and a protocol for culturing a dangerous human pathogen; suicide instructions were blocked. Co-founder Devon (last name withheld; still employed elsewhere) says the company is revenue-funded, talking to VCs, has cloud-provider deals, and no KYC beyond a credit card. CivAI’s Andrew Yoon called abliterated models “a sociopath.” Defenders are split: some say they need the same tools as attackers, others prefer fine-tunes and note abliteration can also strip capability.