rotative
the decoder · September 6, 2026

Abliteration.ai sells stripped-down AI models for $5/m token

abliteration.ai removes safety guardrails from open-weight models like glm-5.3 and sells access for $5/m token, targeting cybersecurity testing but risking misuse.

read in 简体中文 · 繁體中文 · 日本語 · 한국어

abliteration.ai, a us startup, offers a commercial api for modified versions of open-weight ai models like glm-5.3, which have had their safety guardrails removed. the company uses a process called abliteration to tweak the model weights, suppressing internal activation patterns that trigger refusals. the modified glm-5.3 model, called abliterated-model-large-v2, is marketed for offensive cybersecurity, ai red teaming, and trust and safety work. customers can access the model without running their own gpu infrastructure, but this ease of access also increases the risk of misuse. the service costs $5 per million input or output tokens.

read at source ↗

← rotative