AI Coding Live News

TechCrunch reports startup Abliteration.ai turned the long-standing open-source abliteration technique (orthogonalization that removes refusal tendencies) into a hosted browser/API service, including an abliterated build of Z.ai's recent open-weight GLM-5.3. The company says it targets "offensive cyber, red-teaming, and agent testing work other models refuse to do." TechCrunch created a free account and says the abliterated GLM-5.3 complied with requests for a Chrome password-stealer script and a protocol for culturing a dangerous human pathogen at home. The platform offers a customer-side moderation layer and keeps some limits (e.g. suicide instructions in testing) but has no full KYC beyond logging the payment card. Safety researchers (including CivAI) warn scaled access could enable harm; some red-teamers say they still rely more on fine-tuning open models and that abliteration can reduce capability.

Mistral AI introduced Shieldstral, a 3B open-weights content-safety model that can be deployed on-device. It takes a moderation policy as a plain-language question and returns a calibrated score, with one interface for text and images. A technical report is up at arXiv:2607.25857.

Mistral launched Shieldstral, a 3B Apache 2.0 open-weights content safety model optimized for on-device deployment on a single 16GB GPU with natural-language moderation policy evaluation.