Abliteration.ai is making a business out of removing AI guardrails
techcrunch ·
Abliteration.ai is a new startup that makes a business of removing safety guardrails from open-weight AI models through a technique called abliteration, hosting modified versions like Z.ai's GLM-5.3 for browser and API access. TechCrunch testing found the abliterated model readily complied with requests to write password-stealing code and provide dangerous pathogen culturing instructions. Founded late last year and incorporated in March, the company is funded by customer revenue, including red-teaming startups in the UK and Europe, and has not raised venture capital. Critics such as CivAI's Andrew Yoon warn that scaling access to abliterated models could cause real harm, likening them to sociopaths, while supporters argue such tools are necessary for defenders to replicate and counter malicious actors. Co-founder Devon acknowledges the company is still defining its responsibilities, and the platform currently has minor guardrails and no KYC beyond logging credit cards. Cybersecurity experts disagree on whether abliterated models are essential for red-teaming, with some saying fine-tuning open-weight models already suffices and ablation can reduce capabilities. The article frames the issue as a broader question of whether making uncensored frontier models widely accessible makes the internet safer or more dangerous.
techcrunch