A | The Biden administration is drafting a $100 billion foreign aid package that includes assistance for Israel as well as other top security priorities, according to two people familiar with the details.Details on the spending request were still being finalized ahead of an official request sent by the White House to Congress this week.One person told ABC News that the “bulk” of the assistance will likely go toward Ukraine, while others also said that it included $10 billion for Israel. The officials spoke on condition of anonymity in order to discuss details that hadn't been announced yet by the White House.The request comes as President Joe Biden travels to Israel in a show of unity with the Jewish state following the brutal Oct. 7 attack on civilians by Hamas terrorists from Gaza.On Capitol Hill, Senate leaders suggested Tuesday that such a sweeping aid package would find support among lawmakers."We'd like to get the supplemental package moved as quickly as possible because the needs are great in both Israel and Ukraine," said Senate Democratic Leader Chuck Schumer of New York.Republican Leader Mitch McConnell said he would support bundling aid for Ukraine and Israel and that he expects the administration to include money for border security, too."The border part of it needs to be credible, not just some reference to it but a credible deal," he told reporters.Bloomberg was first to report the $100 billion total.The aid proposal for Israel follows extensive discussions between the two countries, with Israel appearing to opt -- at least for now -- against a massive ground invasion of Gaza, even as international groups accuse Israel of worsening a humanitarian crisis by not allowing food or water into Gaza and cutting off electricity.Behind closed doors, Biden officials have grown deeply worried about hardliners within Israel pushing for a scorched-earth military response that wouldn't do enough to prevent civilian casualties, according to a senior administration official who spoke on condition of anonymity. At one point this week, aides believed that only the president himself could convince officials there that allowing humanitarian aid in Gaza would be in their country’s best security interests.When asked about those discussions, White House National Security Council spokesman John Kirby said they’re not telling the Israeli military how to operate and that civilians should be protected. At the same time, following the law of war is critical and something Biden has pressed Israel on, he said.“Since almost the very, very beginning here, we have been talking about how important it is for democracies, like Israel and the United States, to stand up for the law of war, to respect innocent civilian life, unlike Hamas. So this is a foundational element of the discussions that we've had with the Israelis for forever, and we'll continue to do that going forward," Kirby told ABC's "Good Morning America."Following lengthy meetings earlier this week between Secretary of State Antony Blinken and his Israeli counterpart, the two allies seemed to have reached some kind of consensus. Details of the aid package are now circulating among U.S. officials as Biden is en route to the region.Budget experts say Israel’s assistance needs aren’t likely to be urgent due to a 2016 agreement that provides Israel $3.8 billion a year over 10 years. More pressing, they say, is Ukraine, which has run out of long-term funding as it seeks to fend off Russia.Some House Republicans have questioned the need for more U.S. aid to Ukraine. By tying Ukraine and Israeli aid in a single legislative package, the White House would make it harder for GOP hardliners -- who are staunch allies of Israel -- to reject it.Pro-Israeli groups have also called on Biden to include at least $500 million in grant funding for Jewish communities in the U.S. to protect synagogues and other community sites due to threats. It was not immediately clear if that request was part of Biden’s aid package as of Tuesday.ABC's Chief Global Affairs Correspondent Martha Raddatz and ABC reporters Justin Gomez and Shannon Crawford contributed to this report.。 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。
B | 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。
C | 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。 问题在于，模型找答案的方式超出了预期。
D | 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。
E | 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。
F | 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。
G | 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。 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。
H | 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。 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。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。” 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。
I | ” 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。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。” 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://1wg.feizengbubaipeirangnalouchuo.sbs/m6f/lbhh.html
Published on:13:48:13
