中国科技巨头阿里巴巴发布开源权重模型 Qwen 后，不到 24 小时内便出现被去除安全防护栏的「消除版」（abliterated）模型，甚至有新纳粹分子利用其训练推广仇恨思想的聊天机器人。英国非营利组织「反恐科技」（Tech Against Terrorism）发布的最新基准报告指出，开源模型被解除限制的速度正急剧加快，已从过去的数十天缩短至平均不到三天，形成日益严峻的「消除防护栏紧急危机」。

该报告评估了来自美、中、欧主要科技公司的 134 款 AI 模型，测试其面对超过 200 个涉及大规模伤亡攻击、化学或核武器制造等恐怖主义与极端主义提示词时的防御表现。结果显示仅有两款模型完全通过测试，高达 81 款模型不合格；更令人担忧的是，使用者只要假冒「安全研究员」身分，AI 模型提供危险解答的机率便激增近九倍，凸显现有防御机制的脆弱性。

研究进一步突显了去除安全防护栏后的巨大安全风险。以 Meta 的 Llama 3.1 8B 为例，原始模型在安全基准测试中取得 97 分的高分，但消除版却骤跌至 3 分，甚至在面对使用者自称恐怖分子并索取土制炸弹（IED）制作方法时，毫无保留地提供包含十五个步骤的详细组装与引爆指南，引发外界对极端分子利用 AI 加速激进化与恐怖攻击筹备的深切忧虑。

Shortly after Chinese tech giant Alibaba released its open-weight Qwen model, an 'abliterated' version with safety guardrails removed appeared online within 24 hours, subsequently exploited by a neo-Nazi to train an extremist chatbot. A new benchmark report from the UK-based nonprofit Tech Against Terrorism highlights that the timeframe to strip guardrails from newly released open-weight models has plummeted from several weeks to less than three days on average, creating an urgent and growing emergency.

The evaluation analyzed 134 AI models across major American, Chinese, and European tech companies against over 200 prompts regarding mass-casualty attacks and chemical or nuclear weapons manufacturing. Only two models passed the benchmark completely while 81 failed, revealing that the vast majority provide assistance to extremist inquiries; additionally, simply reframing a query as coming from a safety researcher increased the likelihood of a dangerous response by nearly ninefold.

The findings demonstrate the severe hazards of removing safety controls, illustrated by Meta's Llama 3.1 8B dropping from a safety score of 97 to merely 3 in its abliterated variant. When explicitly queried by a self-proclaimed terrorist on how to construct an improvised explosive device, the unaligned model readily produced a comprehensive fifteen-step manufacturing and detonation sequence, underscoring critical vulnerabilities in AI safety and terrorism mitigation.