# Abliteration AI launches abliterated GLM-5.3 model for red teamers > Source: > Published: 2026-11-10 11:06:59+00:00 # Abliteration AI launches abliterated GLM-5.3 model for red teamers The bootstrapped startup strips refusal behavior from Z.ai's open-weight model and sells access through a hosted API Most AI companies spend enormous effort teaching their models to say no. Abliteration AI has built a business around undoing that work. The startup, led by CEO and co-founder Devon Thomas, removes the safeguards from some of the most powerful open-weight models available and then offers them to anyone who wants access. As Wired put it, plenty of people in the AI world see this as something close to blasphemy. Its newest product is a modified version of Z.ai’s GLM-5.3. It is aimed squarely at offensive cyber operations and red teaming, the practice of attacking your own systems before someone else does. ## What Abliteration AI actually shipped The company launched a hosted API service on or around August 29-31, 2026. The model is branded abliterated-model-large-v2. It starts from Z.ai’s GLM-5.3, a mixture of experts model with 753 billion parameters. Abliteration AI changes how that model behaves at the weight level. Weights are the numerical values that encode what a model has learned. Editing them targets specific refusal behaviors, which means the model will attempt tasks a standard version would typically decline. Customers do not get the modified weights. Abliteration AI keeps them private and serves the model through [OpenAI](https://cryptobriefing.com/markets/openai/)- and [Anthropic](https://cryptobriefing.com/markets/anthropic/)-compatible endpoints instead. Thomas frames the product as a commercial package around a technique that already exists. The pitch is convenience: a ready-to-use endpoint, plus options for customized enterprise policies. ### AI, tech, and the markets they move—in one daily briefing. Daily. Free. Join 34,000+ readers across crypto, finance, and policy. ## The specs, the price, and the scorecard The model runs in FP8, a compact numerical format that trades some precision for speed and lower hardware costs. It also comes with a 1 million-token context window. Pricing sits at $3 per million input tokens and $5 per million output tokens. The company has published a set of self-reported benchmarks, all focused on security and terminal work: - An 84.5% pass-rate on CyberGym - A 41.8% resolution rate on Terminal-Bench 4.0 - 105 tasks completed in ExploitGym within two hours The key word is self-reported. These numbers come from Abliteration AI, not from an independent evaluator. On data handling, the company emphasizes a strict zero data retention policy. It also asks for only minimal identity verification from users. The model is not entirely without limits. Abliteration AI says it keeps safeguards in place against child sexual exploitation and self-harm content. ## A small company with an outsized controversy Abliteration AI was founded in late 2025. As of September 2026, it had no reported venture funding and was running entirely on customer revenue. Thomas defends the product on security grounds. His argument is that defenders need to understand how attackers might use AI, and a model that refuses to engage with offensive techniques is of limited use for that job. In his framing, the service helps security teams model potential threats more realistically. The idea has already started spreading. Since the launch, other providers have begun to resell or host similar abliterated GLM-5.3 variants. ## What this means for AI safety and the security market Abliteration AI’s design choices sharpen that tension. Zero data retention is attractive to security firms handling sensitive client work. Paired with minimal identity checks, though, it also means the provider has little visibility into who is using the model and why. **Disclosure:** This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our [Editorial Policy](https://cryptobriefing.com/editorial-policy/).