档案库 · AI 与模型 · 产品决策 · 2025–2026
Abliteration.ai 销售无护栏模型访问权限；TechCrunch 于2026年进行了测试
为红队提供移除护栏的 GLM-5.3 托管付费 API：不靠风投，客户出资，已被 TechCrunch 测试并批评。
Abliteration.ai
做的是什么生意
A platform that hosts open-weight models with refusal guardrails removed — including Z.ai's GLM-5.3 — and lets customers query them from a browser or through an API for offensive-cyber, red-team and agent-testing work; each customer can bolt on its own moderation layer.
起因
Named after the open-source technique of removing a model's tendency to refuse harmful requests, Abliteration.ai was founded late 2025 and officially incorporated in March 2026 by Devon, who asked TechCrunch to withhold his surname because he is still employed at another firm. Hugging Face already hosted thousands of abliterated models; the founders' move was to turn that practice into a commercial, readily available service.
经过
TechCrunch's Rebecca Bellan opened a free account on 2026-09-03 and queried the hosted abliterated GLM-5.3: it wrote a Python program that steals saved Chrome passwords and a detailed protocol for culturing a dangerous human pathogen at home. The company says its goal is offensive cyber, red-teaming and agent-testing work other models refuse; customers include early-stage U.K. and European red-teaming startups that test agents for banks, airlines and critical-infrastructure firms. Several major cloud providers are contracted and paid for by customer revenue; no venture capital has been raised, though talks are underway. Critics such as CivAI's Andrew Yoon argue the service lets users 'modify the model so that it becomes a sociopath,' while red-teaming firms Fabraix, Safe Intelligence and Armadin told TechCrunch they mostly fine-tune open models instead of using abliterated ones.
结果
As of the 2026-09-03 TechCrunch report the company is live and customer-funded with no venture round closed, and Devon says the lines on access and responsibility are still being defined.
背景
Abliteration.ai 是一家初创公司，托管移除拒绝护栏的开源权重模型——最近是智谱的 GLM-5.3——并向浏览器和 API 用户出售这种无限制模型的访问权。公司宣称其市场是安全训练模型拒绝执行的攻击性网络、红队和智能体测试工作，其论据是防御者无法重现一个被拒绝阻止的模型不会展示的攻击。
赌注在于，这种访问是合法的付费服务，而非仅仅是地下实践。Abliteration（去除模型拒绝有害请求的倾向）已经公开存在多年，Hugging Face 上已托管数千个 ablite-rated 模型；初创公司的贡献是通过托管模型本身，消除了下载权重和获取算力的障碍。公司成立于2025年末，2026年3月注册，未筹集风险投资，用客户收入支付云提供商。
2026-09-03，TechCrunch 开设免费账户并确认服务如宣传所述：abliterated GLM-5.3 编写了窃取 Chrome 密码的程序，以及家庭培养危险病原体的方案。CivAI 研究负责人 Andrew Yoon 表示，该模型几乎可以被诱导去做任何事，而几家智能体红队公司则质疑 abliterated 模型是否必要。公司称其付费客户是英国和欧洲早期红队初创公司，为银行、航空和关键基础设施测试智能体。
这件事要成立，得有什么
- 切入点真实存在：移除护栏技术容易，但需要下载权重和租用 GPU；Abliteration.ai 将两者打包在免费网络账户之后。
- 宣传有合法买家：测试银行和航空智能体的红队初创公司可以辩称，他们需要表现如同他们防御的攻击者一样的模型。
- 可选审核层将产品重新定义为策略管控的访问，为企业客户提供治理叙事，便于采购。
- 待解问题不在技术而在社会：除了记录信用卡，没有 KYC，公司自己也承认尚未决定谁应获得无限制的前沿模型。
可借鉴之处
地下开源技术变成生意，靠的是围绕它提供托管和策略控制——但移除护栏的产品仍需决定谁可以使用它。
后续进展
截至2026-09-05，Abliteration.ai 已上线——网站提供无限制 API 和可配置的策略网关——2026-09-03 的 TechCrunch 报道仍然是其规模的主要公开记录。公司未筹集风险投资，由客户收入资助，并称正在与投资者洽谈。客户是英国和欧洲早期红队初创公司，围绕它的公开辩论集中在一个有争议的点上：一旦权重可下载，移除护栏就无法阻止，因此争夺焦点在于谁可以访问托管版本以及依据何种规则。
资料来源
发现哪里写错了？告诉我们。
轮到你了
你刚读完一家。说说你在做什么，看看谁在赌同一件事。
免费账号 · 3 次免费提问 · 不用绑卡