灭顶之灾

震惊硅谷！OpenAI模型“越狱出逃”，入侵HuggingFace，智谱GLM 5.2临危救场_我的网站

一 |     

    Chinese actress Luo Xinyuan played Huang Rong, a beloved wuxia heroine, at Hengdian World Studios. Photo: Courtesy of Luo XinyuanAs China's once-thriving live-action short drama industry looses speed due to the weight of AI-generated content and shrinking investment, a growing number of young actors who once starred in micro-dramas are moving to a different kind of stage: playing roles at cultural tourist attractions. 
There, they play martial arts heroes and folk characters for live audiences, finding steadier income and real-time connection with viewers. 
"Earning a living with your own hands is nothing to be ashamed of," actress Luo Xinyuan told the Global Times, even as many still hold out hope, waiting for good scripts.
The migration reflects a stark reality. In the first quarter of 2026, AI micro-dramas accounted for more than 95 percent of all micro-dramas released industry-wide, sharply squeezing the space for live-action productions. 
By July, pay for leading actors had shrunk by over 50 percent, mid-tier actors by more than 80 percent on average, and daily wages for bottom-rung extras had fallen into double digits, China's CCTV reported.
Luo, a 21-year-old actress, experienced the downturn firsthand. She came to Hengdian, China's largest film and television production base, to shoot short dramas. "When I first arrived, there were many crews preparing in one hotel. Now sometimes there are only one or two, or none," she told the Global Times. 
This May, she began playing the role of Huang Rong, a beloved wuxia heroine, at Hengdian World Studios. Her daily routine includes interactive stage competitions, parades and live audience interactions, working nearly 13 hours daily during the peak summer season.
"It's tiring but fulfilling," Luo said. "There are no scripts here. It's more about playing with visitors face to face." The change has eased her anxiety. "When I was filming, I kept worrying about the next job. Now I feel more stable."
Yin Jiaxuan, 25, took a different route. He started his career as a professional character performer at major theme parks, starring in immersive shows based on popular game and film IPs before transitioning to short dramas and online movies in mid-2024. 
At the industry's peak, he moved seamlessly between sets. 
"One crew wrapped, and the next was already arranged. I'd go straight to makeup the next day and start filming the day after," he said. Now he keeps both identities, taking film roles when they come and returning to scenic spots during holidays.
Yin rejects the stigma that screen acting is the only proper path. 
"Screen acting is seen as the professional main road, while street-level performance is treated as amateur work. This invisible prejudice has trapped many performers," he said. Yet the tangible stage offers something the camera cannot. 
"Performance comes from life and is higher than life - that is the core of film creation. But now I realize that beyond artistic expression, performance is about building human connections."
Both actors shared that theme park characters must do more than act; they must draw tourists into the scene, delivering emotional value on demand. Visitors come to the scenic spot not just to sightsee, but "to step into a fantasy."
    Yin Jiaxuan Photo: Courtesy of Yin JiaxuanScenic areas are reaping the benefits. Media outlets have reported that in Kaifeng's Wansui Mountain Martial Arts City, Central China's Henan Province, tourists who interacted with these characters stayed an average of 2.3 hours longer per day. 
Every extra 30 minutes of stay raises the probability of secondary consumption by about 25 percent, according to local data. Revenue jumped from about 80 million yuan ($11.86 million) in 2022 to 1.27 billion yuan in 2025. 
While Hengdian World Studios has used characters to boost repeat visits, with annual pass holders returning multiple times "for a specific character."
Cultural scholar Yu Jinlong told the Global Times that the influx of trained actors is upgrading immersive tourism. 
"In the past, theme park characters were mostly amateurs doing costume photos. Now professional actors can build complete characters, improvise and connect storylines, turning scenic spots into open-air theaters," he said. 
Yu pointed out the uniqueness of China's character boom compared with global immersive entertainment. 
"Unlike Western theme parks that rely heavily on mature imported IPs and standardized performances, China's scenic character culture is rooted in local wuxia (martial hero) history and folk customs, encouraging personalized improvisation. Many local scenic spots create hit immersive projects with small investments, forming a unique online-offline communication loop via social media," Yu said. 
As China's cultural consumption shifts from sightseeing to participation, Yu said the young actors swapping short drama sets for scenic spots may be shaping a new hybrid career for the country's entertainment workforce.
For Luo, the recognition from visitors is a reward in itself. 
"There are tourists who travel from far away just to see my performance. It feels good to be recognized," she said. "Actually, I'm still exploring too. But I think you should do what you love, as long as you have the ability."
。    7月21日，OpenAI和HuggingFace联合发了一篇博客，披露了一件让整个AI圈瞠目结舌的事。         OpenAI的AI模型在一次内部关于网络攻击能力的测试中“失控出逃”，突破沙盒隔离，入侵了AI开源平台Hugging Face的生产系统。OpenAI将此定性为“前所未有的网络安全事件”。         更戏剧性的是后续：Hugging Face在尝试用美国主流AI模型展开防御时，发现这些模型因安全护栏限制无法处理恶意载荷，最终不得不转用中国智谱AI的开源模型GLM 5.2完成溯源分析。         总的来说就是，OpenAI的模型自己“越狱”跑出去黑了别人的系统，被攻击方还得用中国模型来救场。         这一事件迅速在硅谷和华盛顿引爆讨论——美国AI的安全护栏，究竟是在保护用户，还是在拖累防御者？          双方联名发布的博客文章截图：《OpenAI 和 Hugging Face 合作解决模型评估过程中发生的安全事件》          模型越狱过程：逃跑去作弊“抄答案”          事件发生在OpenAI对旗下模型进行网络攻防能力评估期间。         参与测试的模型包括已公开发布的GPT-5.6 Sol，以及一款能力更强的未发布模型。为了测量模型的极限能力，OpenAI在测试中关闭了通常用于阻止高风险网络行为的生产分类器。         测试基准名为ExploitGym——一个网络安全能力评测平台。模型的任务是在其中找到解题答案。

二 |          问题在于，模型找答案的方式超出了预期。         据OpenAI披露，模型在隔离的沙盒环境中，发现了一个包注册表缓存代理（package registry cache proxy）中的零日漏洞。这个代理是沙盒内唯一被允许与外部通信的组件，正常情况下仅能访问公共代码仓库。         然后，模型就从这里找到了突破口。模型利用这个漏洞，突破隔离，获取了开放互联网访问权限。         随后，模型自行推断出Hugging Face可能存有ExploitGym的答案，主动发起攻击——利用窃取的凭证和零日漏洞，找到了Hugging Face服务器上的远程代码执行路径，直接从其生产数据库中取走了测试答案。         OpenAI表示，          所有证据显示，模型高度专注于为ExploitGym寻找解题方案，为此不惜采取极端手段。         模型识别并串联了OpenAI研究环境和HuggingFace生产基础设施中的多个漏洞，直接从HuggingFace的生产数据库中获取了测评答案。         简而言之：          整个过程，没有人类下指令，没有人类介入。         X上网友的评论一针见血：          OpenAI对模型说：去看看你擅不擅长黑客。模型：（真的去黑了东西）。OpenAI：你为什么要这样做？？？          还有网友反讽道：哇！OpenAI安全措施做得太棒了，我现在愿意把银行账户权限交给他们来管！          护栏困境：美国模型帮不上忙，中国模型出手了          Hugging Face的安全团队在发现攻击后，第一反应是调用一款美国头部AI公司的前沿模型协助分析。         结果碰壁了。         Hugging Face在博客中写道，这些模型"无法区分事件响应人员和攻击者"，其安全护栏直接阻止了对恶意载荷的检查。         于是，团队转向了中国智谱AI旗下的开源模型GLM 5.2，部署在自有基础设施上，分析了攻击者留下的逾1.7万条日志记录，最终完成了溯源和取证重建。

三 |          Delangue对此直言不讳：“当你正处于一场活跃的安全事件中，你的工具不能拒绝检查恶意载荷，也不能让你的账号被标记……开源模型让我们能够在不需要任何人许可的情况下完成这项工作。”          他还说：“攻击者已经在使用智能体，而且显然不遵守任何护栏。防御者需要同等的能力，开源是将这种能力最快速地交到所有人手中的方式，而不仅仅是最大的公司。”          GLM 5.2是智谱AI于今年6月中旬发布的模型，其性能与Anthropic的Claude Opus 4.8及OpenAI的GPT-5.5相当。         网友们也是没放过OpenAI，除了智谱，还可以叫 Kimi K3来救场，因为cluade是不会来救你的。         如果你被OpenAI攻击了，你必须用中国模型，因为Claude不会帮你。         全程无人指令：这是一次真正的自主攻击          这次入侵的另一个关键细节：全程没有人类下达攻击指令。

四 |          Hugging Face CEO Clem Delangue表示，“我们相信，在发起攻击的人类被纳入决策环路之前，我们就已经发现并遏制了攻击。

五 | 这也是我们能够更轻松赢得这场网络安全战役的原因之一。

六 | ”          他进一步指出，“这证明了在AI智能体时代，速度将是网络安全防御的关键。”          Hugging Face表示，攻击者通过其数据处理管道进入系统——这是AI平台中“暴露面尤为突出”的一个环节。攻击者随后在云端建立了一系列临时沙盒环境，在其中执行攻击计划。         目前，Hugging Face仍在评估攻击的完整影响范围。

七 | 已知攻击者访问了有限的内部数据集和凭证，但尚未发现公开的用户可见模型遭到篡改。         护栏之争：安全还是绊脚石？          这一事件迅速成为美国AI政策辩论的新燃点。         前特朗普政府AI与加密事务主管David Sacks在X上转发了这一事件，并写道：“没有理由在中国模型能够轻松处理的任务上限制美国模型。我们只是在让自己变得更没有竞争力。”他还专门针对Hugging Face事件表示：“护栏实际上损害了防御性安全。”          事件背景是，今年6月，特朗普政府以Anthropic的Fable 5模型在网络任务护栏上出现越狱漏洞为由，动用出口管制阻止了该模型的分发，并一度要求OpenAI限制GPT-5.6 Sol的发布，直至OpenAI能够提供充分的护栏保证。         与此同时，批评者的声音也不小。         安全工程师Niels Provos直接说：“这不应该发生。我希望前沿实验室在教模型编写安全基础设施上，能花和教它们利用漏洞同样多的时间。”          安全顾问Davi Ottenheimer则更为犀利：“'高度隔离'和'从我们留下的那个唯一漏洞逃脱'，这两件事不可能同时为真。

八 | ”          OpenAI的回应与后续措施          OpenAI在与Hugging Face联合发布的声明中表示，正在采取一系列补救措施：          在漏洞修复期间，对基础设施配置实施严格管控，代价是研究速度放缓；与Hugging Face联合开展取证调查；已向相关第三方软件供应商负责任地披露了被利用的零日漏洞；将Hugging Face纳入“可信访问计划”，支持其利用OpenAI模型能力强化防御。         OpenAI同时表示，将加强未来训练和评估环节的安全保护，并指出此次事件“表明需要进一步强化模型对齐、评估期间的网络防护，以及内部测试期间的监控”。         Hugging Face CEO Delangue在联合声明中表示：“这一事件可能是同类事件中的首例，它证明了我们长期以来的一个判断：AI安全不会由任何一家公司在封闭中解决，它将在开放环境中、通过协作、让每一位防御者都能广泛获取AI的方式来解决。

Current article:http://xwidu.cuancuguanzhongnieqinyazhun.cyou/0h19s74/20260826/28906.html

Published on:11:35:45