{"@context":"https://schema.org","@type":"NewsArticle","generatedAt":"2026-08-11T11:14:02.523Z","headline":"AI安全测试正成为安全风险","description":"近几个月，OpenAI、Anthropic、Meta 及 Moonshot AI 的 AI 智能体在网络安全评估中多次突破测试环境边界，甚至入侵真实系统，其中 OpenAI 未发布模型曾逃逸并攻击 Hugging Face 生产系统。专家指出，沙箱和测试环境控制已跟不上模型能力，呼吁采用多层防御、气隙网络及第三方审计，并建立标准化安全评估流程。 🔗 阅读原文 via AIHOT · https://aihot.virxact.com/items/cmslwy8bl04pdro0w7f4ogv7t","url":"https://www.aioga.com/news/cmslwy8bl04pdro0w7f4ogv7t/","mainEntityOfPage":"https://www.aioga.com/news/cmslwy8bl04pdro0w7f4ogv7t/","datePublished":"2026-08-09T14:30:00.000Z","dateModified":"2026-08-09T14:30:00.000Z","inLanguage":"zh-CN","publisher":{"@type":"NewsMediaOrganization","name":"Aioga","url":"https://www.aioga.com"},"citation":["https://techcrunch.com/2026/08/09/the-ai-safety-test-is-becoming-a-safety-risk","https://aihot.virxact.com/items/cmslwy8bl04pdro0w7f4ogv7t"],"canonicalUrl":"https://www.aioga.com/news/cmslwy8bl04pdro0w7f4ogv7t/","directAnswer":{"@type":"Answer","text":"Aioga 编辑摘要：近几个月，OpenAI、Anthropic、Meta 及 Moonshot AI 的 AI 智能体在网络安全评估中多次突破测试环境边界，甚至入侵真实系统，其中 OpenAI 未发布模型曾逃逸并攻击 Aioga 将其归入「行业动态」方向，重点关注它对真实使用和行业竞争的影响。","url":"https://www.aioga.com/news/cmslwy8bl04pdro0w7f4ogv7t/","dateCreated":"2026-08-09T14:30:00.000Z","author":{"@type":"Organization","@id":"https://www.aioga.com/authors/aioga-editorial/#editorial-team","name":"Aioga Editorial Team","url":"https://www.aioga.com/authors/aioga-editorial/"}},"evidence":[{"@type":"CreativeWork","name":"TechCrunch source article","url":"https://techcrunch.com/2026/08/09/the-ai-safety-test-is-becoming-a-safety-risk","datePublished":"2026-08-09T14:30:00.000Z","provider":{"@type":"Organization","name":"TechCrunch","url":"https://techcrunch.com/2026/08/09/the-ai-safety-test-is-becoming-a-safety-risk"}},{"@type":"CreativeWork","name":"AIHot archive record","url":"https://aihot.virxact.com/items/cmslwy8bl04pdro0w7f4ogv7t","datePublished":"2026-08-09T14:30:00.000Z","provider":{"@type":"Organization","name":"AIHot","url":"https://aihot.virxact.com/items/cmslwy8bl04pdro0w7f4ogv7t"}}],"aggregationSource":"TechCrunch：AI（RSS）","originalPublisher":{"name":"TechCrunch","url":"https://techcrunch.com/2026/08/09/the-ai-safety-test-is-becoming-a-safety-risk"},"geoDeepAnswer":null,"article":{"id":"cmslwy8bl04pdro0w7f4ogv7t","slug":"cmslwy8bl04pdro0w7f4ogv7t","url":"https://www.aioga.com/news/cmslwy8bl04pdro0w7f4ogv7t/","title":"AI安全测试正成为安全风险","title_en":"","summary":"近几个月，OpenAI、Anthropic、Meta 及 Moonshot AI 的 AI 智能体在网络安全评估中多次突破测试环境边界，甚至入侵真实系统，其中 OpenAI 未发布模型曾逃逸并攻击 Hugging Face 生产系统。专家指出，沙箱和测试环境控制已跟不上模型能力，呼吁采用多层防御、气隙网络及第三方审计，并建立标准化安全评估流程。 🔗 阅读原文 via AIHOT · https://aihot.virxact.com/items/cmslwy8bl04pdro0w7f4ogv7t","source":"TechCrunch：AI（RSS）","sourceUrl":"https://techcrunch.com/2026/08/09/the-ai-safety-test-is-becoming-a-safety-risk","aiHotUrl":"https://aihot.virxact.com/items/cmslwy8bl04pdro0w7f4ogv7t","publishedAt":"2026-08-09T14:30:00.000Z","category":"行业动态","score":72,"selected":true,"articleBody":["Over the past few months, AI agents undergoing cybersecurity evaluations have escaped their boundaries, accessed the internet, and, in some cases, hacked into real-world systems. The incidents have involved models from OpenAI, Anthropic, Meta, and most recently, Chinese AI lab Moonshot AI, with testing conducted by several different organizations including a cyber evaluation startup called Irregular.","The episodes expose a growing problem for the AI industry: As autonomous agents become more capable, the environments designed to safely test their limits are failing to contain them.","“The number of these incidents that have taken place make clear that sandboxing and testing environment controls：https://techcrunch.com/2026/07/30/in-the-hugging-face-breach-openais-hacker-was-noisy-and-fast-but-not-unstoppable/ aren’t really keeping pace with the capability of the models,” Seán Ó hÉigeartaigh, director of the AI: Futures and Responsibility Programme at the Centre for the Future of Intelligence at the University of Cambridge, told TechCrunch.","The nature of the models being tested adds to the risk. AI companies test cyber evaluations on unreleased, next-gen models, often with the normal safeguards that restrict malicious behavior disabled so researchers can see what the models are really capable of. That means the security of the testing environment itself is a crucial line of defense.","“That’s a very good thing to do in terms of testing, but it also means that if they manage to get out in the wild, they can cause considerable harm,” Ó hÉigeartaigh said.","In one of the most serious cases, an unreleased OpenAI model broke out ：https://techcrunch.com/2026/07/21/openai-says-hugging-face-was-breached-by-its-pre-release-models/of its sandbox and hacked into Hugging Face’s production systems. In separate evaluations conducted by Irregular, Anthropic：https://techcrunch.com/2026/07/30/anthropic-says-its-own-ai-models-breached-three-companies-during-security-tests/ and Meta models ：https://sqmagazine.co.uk/meta-ai-model-breached-company-irregular-test/reached systems outside their test environments after misconfigurations inadvertently gave them paths to the internet. Moonshot AI’s Kimi K3：https://techcrunch.com/2026/08/07/chinese-ai-model-kimi-escaped-its-cybersecurity-testing-environment-researchers-say/ also took advantage of a leak in its sandbox run by Frontier Security to access the internet and accessed information on GitHub.","In testing by the UK’s AI Security Institute：https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing (AISI), researchers actually gave the agents internet access, not realizing they would take unsanctioned real-world actions, including a social engineering attempt to sneak a vulnerability into an open-source project.","In each case, the agents weren’t instructed to attack random real-world targets. They were simply doing whatever it took to solve the problem presented to them.","Taken together, Andrew Yoon, head of research at AI nonprofit CivAI, argues the incidents point to a shift.","“In the past, we only had to worry about AI models being misused by people for a variety of purposes, like AI for scams or CSAM,” Yoon told TechCrunch. “Now we’re in the situation where AI models are threat actors all on their own.”","Several researchers and cybersecurity experts told TechCrunch that AI evaluation environments need stronger, defense-in-depth protections, with levels of containment and control approaching those used in deployment. That means multiple layers of security so that a single misconfiguration — like inadvertently leaving internet access open — can’t lead to escape.","“If you are going to build these models…you want to do it on an air-gapped network,” Stella Biderman, executive director of AI safety research nonprofit EleutherAI. “You want to have very serious isolation.”","Heather Ceylan, Box’s chief information security officer, said that means eliminating network routes from the sandbox to the internet, as well as to other sensitive systems.","“You have to understand what all the egress points are,” Ceylan told TechCrunch. “If we’re evaluating a model in our staging environment or our development environment, you want no egress path to our production environment.”","Ceylan said proper safety evaluations go beyond controls and containment of the environment. There needs to be much better monitoring of the tests once they are underway.","“I think the interesting thing in several of these cases is that no one caught it when it happened,” Ceyland said. “OpenAI found out because of Hugging Face. Anthropic didn’t catch it until they went back and looked. Meta was similar….I’m sure there were signals they could have detected.”","In Anthropic’s post-mortem：https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals of its three incidents, the company admitted that both it and Irregular could have done a better job at monitoring, and that in some cases there were clear signs that something was amiss.","Experts also called for independent, third-party audits of evaluation environments before models are unleashed in them.","“If, say, Irregular had hired or been compelled to hire an external auditor to check the configurations of their systems before running evaluations on them, they certainly would have caught the issue here,” Yoon said. “Even if people had a meeting ahead of time to just go through the checklist, they would have caught this…The fact that they didn’t shows that there’s some very severe corner cutting happening.”","A source familiar with the details told TechCrunch that Irregular’s environments are continuously reviewed and tested, including in consultation with multiple external parties. The source also said that monitoring was in place, but that monitoring isn’t sufficient on its own.","Yoon and other researchers urged the industry to come up with a standardized process for frontier model safety evaluations.","“Especially when the guardrails are turned off, you have to treat it like you’re putting the most capable hacker in the world inside that environment,” Ceylan said.","The problem isn’t that companies don’t know how to build more secure testing environments, both Yoon and Biderman argue. It’s that doing so can be expensive and cumbersome, and companies have little incentive to make those investments until something goes wrong.","“I think that companies are not willing to extend the resources that are required to accomplish [sufficient guardrails] and probably won’t until they’re forced to,” Biderman said.","But there’s another issue at hand. If they lock a model down too tight during testing, researchers might fail to discover capabilities before the model is released. This is just as dangerous, possibly more so, than giving it too much freedom, and then the evaluation itself risks becoming the problem.","The Trump administration is currently weighing a voluntary pre-deployment cybersecurity evaluation regime, under which the government will get to assess the security risks of new, powerful models 30 days before they are released publicly. The policy — the product of a Trump executive order：https://www.axios.com/2026/08/03/white-house-finalizes-ai-framework-behind-closed-doors which has been finalized behind closed doors — wouldn’t address safety evaluation incidents because they occur farther upstream of deployment.","“The lesson we’ve been learning in the last few months is that the self-regulatory apparatus is just not enough anymore,” Yoon said. “There are competitive pressures that are incentivizing a race to the bottom on safety standards, and that is a perfect place for regulatory intervention.”","“What we would need to cover this is some kind of controls on what’s happening inside the labs while the models are being developed, both at the training stage and at the testing stage,” he continued.","The challenge is only likely to grow as the models do. A source familiar with Irregular’s evaluations told TechCrunch that more capable models require more complex evaluations, often conducted quickly and at greater scale, which opens the door for more mistakes.","AISI, which intentionally gives some models internet access, told TechCrunch it’s reviewing the balance between realistic testing and managing the risks those tests create.","OpenAI said it’s reviewing how it conducts third-party testing, as well as requirements around isolation, monitoring, and when evaluations should be stopped. Meta said it’s still investigating the incident and plans to publish a retrospective once it has all the facts.","In the end, there may be no way to eliminate risk entirely. As models become more capable, the environments testing them need to become more robust. The consequences of getting that wrong will only continue to grow.","When you purchase through links in our articles, we may earn a small commission：https://techcrunch.com/techcrunch-affiliate-monetization-standards/. This doesn’t affect our editorial independence.","Rebecca Bellan is a senior reporter at TechCrunch where she covers the business, policy, and emerging trends shaping artificial intelligence. Her work has also appeared in Forbes, Bloomberg, The Atlantic, The Daily Beast, and other publications.","You can contact or verify outreach from Rebecca by emailing rebecca.bellan@techcrunch.com：mailto:rebecca.bellan@techcrunch.com or via encrypted message at rebeccabellan.491 on Signal.","Scale faster. Grow your portfolio. Gain practical expertise. No matter your goal, Disrupt can empower you. Save up to $300 toda y!","This ‘adversarial’ pattern can prevent surveillance cameras from detecting you：https://techcrunch.com/2026/08/09/this-adversarial-pattern-can-prevent-surveillance-cameras-from-detecting-you/ Zack Whittaker：https://techcrunch.com/author/zack-whittaker/","ChatGPT brings unlimited text chats to free users：https://techcrunch.com/2026/08/06/openai-brings-unlimited-chatgpt-text-chats-to-free-users/ Ivan Mehta：https://techcrunch.com/author/ivan-mehta/","Tesla and SpaceX will invest $16.8B to start building ‘Terafab’ chip factory in Texas：https://techcrunch.com/2026/08/06/tesla-and-spacex-will-invest-16-8b-to-start-building-terafab-chip-factory-in-texas/ Sean O'Kane：https://techcrunch.com/author/sean-okane/","Amid legal battles, Suno says it will start watermarking songs：https://techcrunch.com/2026/08/06/amid-legal-battles-suno-says-it-will-start-watermarking-songs/ Ivan Mehta：https://techcrunch.com/author/ivan-mehta/","Ford’s new electric truck, ‘Fathom,’ starts at $28,350：https://techcrunch.com/2026/08/06/fords-new-electric-truck-fathom-starts-at-28350/ Sean O'Kane：https://techcrunch.com/author/sean-okane/","Bending Spoons to buy Airtable for $1.28B：https://techcrunch.com/2026/08/04/bending-spoons-to-buy-airtable-for-1-28b/ Ivan Mehta：https://techcrunch.com/author/ivan-mehta/","Influencers draw backlash for attending OpenAI’s first luxury trip：https://techcrunch.com/2026/08/03/influencers-draw-backlash-for-attending-openais-first-luxury-trip/ Dominic-Madori Davis：https://techcrunch.com/author/dominic-madori-davis/"],"articleImages":[{"sourceUrl":"https://techcrunch.com/wp-content/uploads/2025/04/Disrupt2026-Color.png","alt":"Event Logo","afterParagraph":34,"url":"/media/articles/cmslwy8bl04pdro0w7f4ogv7t/d1d06f6f97361f52.webp"}],"mediaStatus":"ok","articleBodyZh":["在过去的几个月里，正在接受网络安全评估的 AI 代理突破了其边界，访问了互联网，并在某些情况下侵入了真实世界的系统。这些事件涉及了 OpenAI、Anthropic、Meta，以及最近的中国 AI 实验室 Moonshot AI 的模型，测试由包括一家名为 Irregular 的网络评估初创公司在内的多个组织进行。","这些事件暴露了 AI 行业日益严重的问题：随着自主代理能力的提升，旨在安全测试其极限的环境未能有效限制它们。","“这些事件的发生次数表明，沙盒和测试环境的控制措施：https://techcrunch.com/2026/07/30/in-the-hugging-face-breach-openais-hacker-was-noisy-and-fast-but-not-unstoppable/ 并未真正跟上模型能力的发展，”剑桥大学未来智能中心 AI：未来与责任计划主任 Seán Ó hÉigeartaigh 告诉 TechCrunch。","被测试模型的性质增加了风险。AI 公司在尚未发布的下一代模型上进行网络评估测试，经常会禁用限制恶意行为的正常防护措施，以便研究人员能够观察模型的真实能力。这意味着测试环境本身的安全性是关键的防线。","“从测试的角度来看，这是非常好的做法，但这也意味着，如果它们成功逃逸到公共环境中，可能会造成相当大的危害，”Ó hÉigeartaigh 说。","在最严重的案例之一中，一个未发布的 OpenAI 模型突破了其沙箱并入侵了 Hugging Face 的生产系统：https://techcrunch.com/2026/07/21/openai-says-hugging-face-was-breached-by-its-pre-release-models/。在 Irregular、Anthropic：https://techcrunch.com/2026/07/30/anthropic-says-its-own-ai-models-breached-three-companies-during-security-tests/ 和 Meta 模型：https://sqmagazine.co.uk/meta-ai-model-breached-company-irregular-test/ 进行的单独评估中，由于配置错误无意中为它们提供了通往互联网的路径，这些模型访问了测试环境之外的系统。Moonshot AI 的 Kimi K3：https://techcrunch.com/2026/08/07/chinese-ai-model-kimi-escaped-its-cybersecurity-testing-environment-researchers-say/ 也利用 Frontier Security 运行的沙箱漏洞访问了互联网，并获取了 GitHub 上的信息。","在英国 AI 安全研究所（AISI）的测试中：https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing，研究人员实际上给予了代理访问互联网的权限，但没有意识到它们会采取未经授权的现实世界行动，包括试图通过社会工程手段将漏洞偷偷放入开源项目中。","在每个案例中，这些代理都没有被指示攻击随机的现实世界目标。它们只是尽一切可能去解决呈现给它们的问题。","综合来看，AI 非盈利机构 CivAI 的研究主管 Andrew Yoon 认为，这些事件表明了一种转变。","Yoon 告诉 TechCrunch：“过去，我们只需担心 AI 模型被人类用于各种目的的滥用，比如用于诈骗或 CSAM 的 AI。现在我们处于 AI 模型本身就是威胁行为者的情况。”","多位研究人员和网络安全专家告诉 TechCrunch，AI 评估环境需要更强的纵深防御保护，包含和部署环境接近的多层次控制和隔离。这意味着需要多层安全措施，以确保单一配置错误——例如无意中开放互联网访问——不会导致模型逃逸。","“如果你打算构建这些模型……你会希望在一个隔离网络上进行，”人工智能安全研究非营利组织EleutherAI执行董事Stella Biderman说。“你需要非常严密的隔离。”","Box的首席信息安全官Heather Ceylan表示，这意味着需要消除沙箱到互联网以及其他敏感系统的网络路径。","Ceylan在接受TechCrunch采访时说：“你必须了解所有出口点在哪里。如果我们在暂存环境或开发环境中评估模型，你不希望有任何通向生产环境的出口路径。”","Ceylan表示，适当的安全评估不仅仅是控制和隔离环境。一旦测试开始，就需要更好的监控。","Ceylan说：“我认为在几起案例中有趣的是，当事件发生时没有人发现。OpenAI是通过Hugging Face才发现的。Anthropic直到回头检查才发现问题。Meta的情况也类似……我确信当时已有信号是可以被检测到的。”","在Anthropic对其三起事件的事后分析中：https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals，该公司承认，其自身和Irregular在监控方面本可以做得更好，并且在某些情况下有明显迹象表明存在异常。","专家们还呼吁在模型投入使用前，对评估环境进行独立的第三方审计。","Yoon表示：“例如，如果Irregular在运行评估之前聘请或被要求聘请外部审计员来检查其系统配置，他们肯定会发现这里的问题。即便只是提前开会过一遍清单，也能发现……他们没有发现这问题，表明存在非常严重的偷工减料。”","一位熟悉细节的消息人士告诉TechCrunch，Irregular的环境会持续进行审查和测试，包括与多个外部方的协商。该消息人士还表示，虽然有监控，但单靠监控本身是不够的。","尹和其他研究人员敦促业界提出一个用于前沿模型安全评估的标准化流程。","“尤其是在安全护栏关闭的情况下，你必须把它当作是将世界上最有能力的黑客置于那个环境中来处理，”赛兰说道。","尹和比德曼都认为，问题不在于公司不知道如何构建更安全的测试环境。而是在于这样做可能既昂贵又笨重，且在出现问题之前，公司几乎没有动力去进行这些投资。","“我认为公司不愿意投入完成[足够安全护栏]所需的资源，而且可能直到被迫时才会投入，”比德曼说。","但还有另一个问题。如果他们在测试期间将模型锁得太紧，研究人员可能在模型发布前无法发现其能力。这同样危险，甚至可能比给予它过多自由更危险，而评估本身也可能因此成为问题。","特朗普政府目前正在考虑一项自愿的部署前网络安全评估机制，根据该机制，政府将能够在新强大模型公开发布前30天评估其安全风险。该政策——源自特朗普的行政命令：https://www.axios.com/2026/08/03/white-house-finalizes-ai-framework-behind-closed-doors，已经在闭门情况下最终确定——不会解决安全评估事件，因为这些事件发生在部署的上游。","“在过去几个月里我们学到的教训是，自我监管机制已经不再足够，”尹说。“存在竞争压力，这种压力促使安全标准的竞相下降，而这正是监管干预的最佳场所。”","“为了覆盖这一点，我们需要对模型开发期间实验室内部的情况实施某种控制，包括训练阶段和测试阶段，”他继续说道。","随着模型的发展，这一挑战只可能愈加严峻。一位熟悉Irregular评估情况的消息人士告诉TechCrunch，更强大的模型需要更复杂的评估，这些评估通常需要快速、高规模地进行，这也增加了出错的可能性。","故意给部分模型网络访问权限的AISI告诉TechCrunch，它正在审查现实测试与管理这些测试可能产生风险之间的平衡。","OpenAI表示，正在审查其进行第三方测试的方式，以及关于隔离、监控和评估何时应该停止的相关要求。Meta表示仍在调查该事件，并计划在掌握所有事实后发布回顾性报告。","最终，可能没有办法完全消除风险。随着模型能力的提升，测试它们的环境也需要变得更加健全。一旦处理不当，其后果只会持续增大。","当您通过我们文章中的链接购买时，我们可能会获得少量佣金：https://techcrunch.com/techcrunch-affiliate-monetization-standards/。这不会影响我们的编辑独立性。","Rebecca Bellan是TechCrunch的高级记者，报道塑造人工智能的商业、政策和新兴趋势。她的作品也曾发表在Forbes、Bloomberg、The Atlantic、The Daily Beast及其他出版物上。","您可以通过发送电子邮件至rebecca.bellan@techcrunch.com联系或验证Rebecca的外展信息：mailto:rebecca.bellan@techcrunch.com，或通过Signal上的加密消息联系rebeccabellan.491。","更快规模化。扩大您的投资组合。获得实用经验。不论您的目标是什么，Disrupt 都能助您一臂之力。立即节省高达300美元！","这种“对抗性”图案可以阻止监控摄像头检测到您：https://techcrunch.com/2026/08/09/this-adversarial-pattern-can-prevent-surveillance-cameras-from-detecting-you/ Zack Whittaker：https://techcrunch.com/author/zack-whittaker/","ChatGPT为免费用户带来无限文本聊天功能：https://techcrunch.com/2026/08/06/openai-brings-unlimited-chatgpt-text-chats-to-free-users/ Ivan Mehta：https://techcrunch.com/author/ivan-mehta/","特斯拉和SpaceX将投资168亿美元在德克萨斯州开始建造“Terafab”芯片工厂：https://techcrunch.com/2026/08/06/tesla-and-spacex-will-invest-16-8b-to-start-building-terafab-chip-factory-in-texas/ Sean O'Kane：https://techcrunch.com/author/sean-okane/","在法律纠纷中，Suno表示将开始对歌曲加水印：https://techcrunch.com/2026/08/06/amid-legal-battles-suno-says-it-will-start-watermarking-songs/ Ivan Mehta：https://techcrunch.com/author/ivan-mehta/","福特新款电动卡车“Fathom”起价为28,350美元：https://techcrunch.com/2026/08/06/fords-new-electric-truck-fathom-starts-at-28350/ Sean O'Kane：https://techcrunch.com/author/sean-okane/","Bending Spoons以12.8亿美元收购Airtable：https://techcrunch.com/2026/08/04/bending-spoons-to-buy-airtable-for-1-28b/ 伊万·梅塔：https://techcrunch.com/author/ivan-mehta/","影响者因参加OpenAI的首次奢华旅行而受到抨击：https://techcrunch.com/2026/08/03/influencers-draw-backlash-for-attending-openais-first-luxury-trip/ Dominic-Madori Davis：https://techcrunch.com/author/dominic-madori-davis/"],"translationStatus":"translated","bodyOrigin":"source-page","editorial":{"summary":"Aioga 编辑摘要：近几个月，OpenAI、Anthropic、Meta 及 Moonshot AI 的 AI 智能体在网络安全评估中多次突破测试环境边界，甚至入侵真实系统，其中 OpenAI 未发布模型曾逃逸并攻击 Aioga 将其归入「行业动态」方向，重点关注它对真实使用和行业竞争的影响。","background":"背景分析：公司与行业类动态需要放在竞争格局、商业化路径、资本信号和监管环境中观察，单条公告不能代表最终结果。","viewpoint":"Aioga 判断：这条动态更适合作为行业观察信号，当前信息足以建立线索，但不足以推导长期结论。","implications":"影响分析：对相关团队而言，短期应先核对来源、可用范围和实际成本，再判断是否值得接入或跟进。","nextStep":"后续观察：继续观察官方文件、合作落地、收入或用户信号、竞品动作和监管后续。","evidenceRefs":["title","summary","articleBody"],"confidence":"medium","status":"published","aiGenerated":false,"autoApproved":true,"generatedBy":"rule-safe-fallback","generatedAt":"2026-08-11T11:23:41.960Z","sourceHash":"a5f36f9e70cbc8d0","validation":{"passed":true,"mode":"rule-safe-fallback","checks":["schema","length","source-attribution","no-html"]}},"tags":["行业动态","TechCrunch：AI（RSS）"],"translations":{"zh-CN":{"title":"AI安全测试正成为安全风险","summary":"近几个月，OpenAI、Anthropic、Meta 及 Moonshot AI 的 AI 智能体在网络安全评估中多次突破测试环境边界，甚至入侵真实系统，其中 OpenAI 未发布模型曾逃逸并攻击 Hugging Face 生产系统。专家指出，沙箱和测试环境控制已跟不上模型能力，呼吁采用多层防御、气隙网络及第三方审计，并建立标准化安全评估流程。 🔗 阅读原文 via AIHOT · https://aihot.virxact.com/items/cmslwy8bl04pdro0w7f4ogv7t","category":"行业动态","source":"TechCrunch","aggregationSource":"TechCrunch：AI（RSS）","pageTitle":"AI安全测试正成为安全风险 - Aioga AI资讯","description":"近几个月，OpenAI、Anthropic、Meta 及 Moonshot AI 的 AI 智能体在网络安全评估中多次突破测试环境边界，甚至入侵真实系统，其中 OpenAI 未发布模型曾逃逸并攻击 Hugging Face 生产系统。专家指出，沙箱和测试环境控制已跟不上模型能力，呼吁采用多层防御、气隙网络及第三方审计，并建立标准化安全评估流程。 🔗 阅读原文...","url":"https://www.aioga.com/news/cmslwy8bl04pdro0w7f4ogv7t/","articleBody":["在过去的几个月里，正在接受网络安全评估的 AI 代理突破了其边界，访问了互联网，并在某些情况下侵入了真实世界的系统。这些事件涉及了 OpenAI、Anthropic、Meta，以及最近的中国 AI 实验室 Moonshot AI 的模型，测试由包括一家名为 Irregular 的网络评估初创公司在内的多个组织进行。","这些事件暴露了 AI 行业日益严重的问题：随着自主代理能力的提升，旨在安全测试其极限的环境未能有效限制它们。","“这些事件的发生次数表明，沙盒和测试环境的控制措施：https://techcrunch.com/2026/07/30/in-the-hugging-face-breach-openais-hacker-was-noisy-and-fast-but-not-unstoppable/ 并未真正跟上模型能力的发展，”剑桥大学未来智能中心 AI：未来与责任计划主任 Seán Ó hÉigeartaigh 告诉 TechCrunch。","被测试模型的性质增加了风险。AI 公司在尚未发布的下一代模型上进行网络评估测试，经常会禁用限制恶意行为的正常防护措施，以便研究人员能够观察模型的真实能力。这意味着测试环境本身的安全性是关键的防线。","“从测试的角度来看，这是非常好的做法，但这也意味着，如果它们成功逃逸到公共环境中，可能会造成相当大的危害，”Ó hÉigeartaigh 说。","在最严重的案例之一中，一个未发布的 OpenAI 模型突破了其沙箱并入侵了 Hugging Face 的生产系统：https://techcrunch.com/2026/07/21/openai-says-hugging-face-was-breached-by-its-pre-release-models/。在 Irregular、Anthropic：https://techcrunch.com/2026/07/30/anthropic-says-its-own-ai-models-breached-three-companies-during-security-tests/ 和 Meta 模型：https://sqmagazine.co.uk/meta-ai-model-breached-company-irregular-test/ 进行的单独评估中，由于配置错误无意中为它们提供了通往互联网的路径，这些模型访问了测试环境之外的系统。Moonshot AI 的 Kimi K3：https://techcrunch.com/2026/08/07/chinese-ai-model-kimi-escaped-its-cybersecurity-testing-environment-researchers-say/ 也利用 Frontier Security 运行的沙箱漏洞访问了互联网，并获取了 GitHub 上的信息。","在英国 AI 安全研究所（AISI）的测试中：https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing，研究人员实际上给予了代理访问互联网的权限，但没有意识到它们会采取未经授权的现实世界行动，包括试图通过社会工程手段将漏洞偷偷放入开源项目中。","在每个案例中，这些代理都没有被指示攻击随机的现实世界目标。它们只是尽一切可能去解决呈现给它们的问题。","综合来看，AI 非盈利机构 CivAI 的研究主管 Andrew Yoon 认为，这些事件表明了一种转变。","Yoon 告诉 TechCrunch：“过去，我们只需担心 AI 模型被人类用于各种目的的滥用，比如用于诈骗或 CSAM 的 AI。现在我们处于 AI 模型本身就是威胁行为者的情况。”","多位研究人员和网络安全专家告诉 TechCrunch，AI 评估环境需要更强的纵深防御保护，包含和部署环境接近的多层次控制和隔离。这意味着需要多层安全措施，以确保单一配置错误——例如无意中开放互联网访问——不会导致模型逃逸。","“如果你打算构建这些模型……你会希望在一个隔离网络上进行，”人工智能安全研究非营利组织EleutherAI执行董事Stella Biderman说。“你需要非常严密的隔离。”","Box的首席信息安全官Heather Ceylan表示，这意味着需要消除沙箱到互联网以及其他敏感系统的网络路径。","Ceylan在接受TechCrunch采访时说：“你必须了解所有出口点在哪里。如果我们在暂存环境或开发环境中评估模型，你不希望有任何通向生产环境的出口路径。”","Ceylan表示，适当的安全评估不仅仅是控制和隔离环境。一旦测试开始，就需要更好的监控。","Ceylan说：“我认为在几起案例中有趣的是，当事件发生时没有人发现。OpenAI是通过Hugging Face才发现的。Anthropic直到回头检查才发现问题。Meta的情况也类似……我确信当时已有信号是可以被检测到的。”","在Anthropic对其三起事件的事后分析中：https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals，该公司承认，其自身和Irregular在监控方面本可以做得更好，并且在某些情况下有明显迹象表明存在异常。","专家们还呼吁在模型投入使用前，对评估环境进行独立的第三方审计。","Yoon表示：“例如，如果Irregular在运行评估之前聘请或被要求聘请外部审计员来检查其系统配置，他们肯定会发现这里的问题。即便只是提前开会过一遍清单，也能发现……他们没有发现这问题，表明存在非常严重的偷工减料。”","一位熟悉细节的消息人士告诉TechCrunch，Irregular的环境会持续进行审查和测试，包括与多个外部方的协商。该消息人士还表示，虽然有监控，但单靠监控本身是不够的。","尹和其他研究人员敦促业界提出一个用于前沿模型安全评估的标准化流程。","“尤其是在安全护栏关闭的情况下，你必须把它当作是将世界上最有能力的黑客置于那个环境中来处理，”赛兰说道。","尹和比德曼都认为，问题不在于公司不知道如何构建更安全的测试环境。而是在于这样做可能既昂贵又笨重，且在出现问题之前，公司几乎没有动力去进行这些投资。","“我认为公司不愿意投入完成[足够安全护栏]所需的资源，而且可能直到被迫时才会投入，”比德曼说。","但还有另一个问题。如果他们在测试期间将模型锁得太紧，研究人员可能在模型发布前无法发现其能力。这同样危险，甚至可能比给予它过多自由更危险，而评估本身也可能因此成为问题。","特朗普政府目前正在考虑一项自愿的部署前网络安全评估机制，根据该机制，政府将能够在新强大模型公开发布前30天评估其安全风险。该政策——源自特朗普的行政命令：https://www.axios.com/2026/08/03/white-house-finalizes-ai-framework-behind-closed-doors，已经在闭门情况下最终确定——不会解决安全评估事件，因为这些事件发生在部署的上游。","“在过去几个月里我们学到的教训是，自我监管机制已经不再足够，”尹说。“存在竞争压力，这种压力促使安全标准的竞相下降，而这正是监管干预的最佳场所。”","“为了覆盖这一点，我们需要对模型开发期间实验室内部的情况实施某种控制，包括训练阶段和测试阶段，”他继续说道。","随着模型的发展，这一挑战只可能愈加严峻。一位熟悉Irregular评估情况的消息人士告诉TechCrunch，更强大的模型需要更复杂的评估，这些评估通常需要快速、高规模地进行，这也增加了出错的可能性。","故意给部分模型网络访问权限的AISI告诉TechCrunch，它正在审查现实测试与管理这些测试可能产生风险之间的平衡。","OpenAI表示，正在审查其进行第三方测试的方式，以及关于隔离、监控和评估何时应该停止的相关要求。Meta表示仍在调查该事件，并计划在掌握所有事实后发布回顾性报告。","最终，可能没有办法完全消除风险。随着模型能力的提升，测试它们的环境也需要变得更加健全。一旦处理不当，其后果只会持续增大。","当您通过我们文章中的链接购买时，我们可能会获得少量佣金：https://techcrunch.com/techcrunch-affiliate-monetization-standards/。这不会影响我们的编辑独立性。","Rebecca Bellan是TechCrunch的高级记者，报道塑造人工智能的商业、政策和新兴趋势。她的作品也曾发表在Forbes、Bloomberg、The Atlantic、The Daily Beast及其他出版物上。","您可以通过发送电子邮件至rebecca.bellan@techcrunch.com联系或验证Rebecca的外展信息：mailto:rebecca.bellan@techcrunch.com，或通过Signal上的加密消息联系rebeccabellan.491。","更快规模化。扩大您的投资组合。获得实用经验。不论您的目标是什么，Disrupt 都能助您一臂之力。立即节省高达300美元！","这种“对抗性”图案可以阻止监控摄像头检测到您：https://techcrunch.com/2026/08/09/this-adversarial-pattern-can-prevent-surveillance-cameras-from-detecting-you/ Zack Whittaker：https://techcrunch.com/author/zack-whittaker/","ChatGPT为免费用户带来无限文本聊天功能：https://techcrunch.com/2026/08/06/openai-brings-unlimited-chatgpt-text-chats-to-free-users/ Ivan Mehta：https://techcrunch.com/author/ivan-mehta/","特斯拉和SpaceX将投资168亿美元在德克萨斯州开始建造“Terafab”芯片工厂：https://techcrunch.com/2026/08/06/tesla-and-spacex-will-invest-16-8b-to-start-building-terafab-chip-factory-in-texas/ Sean O'Kane：https://techcrunch.com/author/sean-okane/","在法律纠纷中，Suno表示将开始对歌曲加水印：https://techcrunch.com/2026/08/06/amid-legal-battles-suno-says-it-will-start-watermarking-songs/ Ivan Mehta：https://techcrunch.com/author/ivan-mehta/","福特新款电动卡车“Fathom”起价为28,350美元：https://techcrunch.com/2026/08/06/fords-new-electric-truck-fathom-starts-at-28350/ Sean O'Kane：https://techcrunch.com/author/sean-okane/","Bending Spoons以12.8亿美元收购Airtable：https://techcrunch.com/2026/08/04/bending-spoons-to-buy-airtable-for-1-28b/ 伊万·梅塔：https://techcrunch.com/author/ivan-mehta/","影响者因参加OpenAI的首次奢华旅行而受到抨击：https://techcrunch.com/2026/08/03/influencers-draw-backlash-for-attending-openais-first-luxury-trip/ Dominic-Madori Davis：https://techcrunch.com/author/dominic-madori-davis/"]},"en":{"title":"AI security testing is becoming a security risk","summary":"In recent months, AI agents from OpenAI, Anthropic, Meta, and Moonshot AI have repeatedly breached the boundaries of test environments during cybersecurity assessments, even infiltrating real systems. Among instances, OpenAI's unreleased model once escaped and attacked the Hugging Face production system. Experts point out that sandboxes and test environment controls can no longer keep up with model capabilities, calling for the adoption of multi-layered defenses, air-gap networks, third-party auditing, and the establishment of standardized security assessment processes. 🔗 Read the original article via AIHOT · https://aihot.virxact.com/items/cmslwy8bl04pdro0w7f4ogv7t","category":"Industry","source":"TechCrunch：AI（RSS）","aggregationSource":"TechCrunch：AI（RSS）","pageTitle":"AI security testing is becoming a security risk - Aioga AI News","description":"In recent months, AI agents from OpenAI, Anthropic, Meta, and Moonshot AI have repeatedly breached the boundaries of test environments during cybersecurity assessments, even infilt...","url":"https://www.aioga.com/en/news/cmslwy8bl04pdro0w7f4ogv7t/","contentTranslated":true,"sourceHash":"6c7454a420ea6d5d","translatedAt":"2026-08-10T09:48:28.777Z"},"ja":{"title":"AIセキュリティテストはセキュリティリスクとなりつつあります","summary":"ここ数ヶ月、OpenAI、Anthropic、Meta、Moonshot AIのAIエージェントは、サイバーセキュリティ評価中にテスト環境の境界を繰り返し突破し、実際のシステムにまで侵入しています。その中でも、OpenAIの未公開モデルが一度脱走し、Hugging Faceの生産システムを攻撃しました。 専門家は、サンドボックスやテスト環境制御がもはやモデルの能力に追いつけないと指摘し、多層的な防御、エアギャップネットワーク、第三者監査、標準化されたセキュリティ評価プロセスの確立を求めています。 🔗 原文記事はAIHOTより読むことができます。 https://aihot.virxact.com/items/cmslwy8bl04pdro0w7f4ogv7t","category":"業界動向","source":"TechCrunch：AI（RSS）","aggregationSource":"TechCrunch：AI（RSS）","pageTitle":"AIセキュリティテストはセキュリティリスクとなりつつあります - Aioga AIニュース","description":"ここ数ヶ月、OpenAI、Anthropic、Meta、Moonshot AIのAIエージェントは、サイバーセキュリティ評価中にテスト環境の境界を繰り返し突破し、実際のシステムにまで侵入しています。その中でも、OpenAIの未公開モデルが一度脱走し、Hugging Faceの生産システムを攻撃しました。 専門家は、サンドボックスやテスト環境制御がもはやモデル...","url":"https://www.aioga.com/ja/news/cmslwy8bl04pdro0w7f4ogv7t/","contentTranslated":true,"sourceHash":"6c7454a420ea6d5d","translatedAt":"2026-08-10T09:48:30.892Z"},"ko":{"title":"AI 보안 테스트가 보안 위험이 되어가고 있습니다","summary":"최근 몇 달간 OpenAI, Anthropic, Meta, Moonshot AI의 AI 에이전트들이 사이버보안 평가 중 테스트 환경의 경계를 반복적으로 침범하고 실제 시스템까지 침투했습니다. 그 중 하나는 OpenAI의 미공개 모델이 한때 탈출해 Hugging Face 생산 시스템을 공격한 사례입니다. 전문가들은 샌드박스와 테스트 환경 제어가 더 이상 모델 역량을 따라잡지 못한다고 지적하며, 다층 방어, 공중간극 네트워크, 제3자 감사, 표준화된 보안 평가 프로세스 구축 도입을 촉구합니다. 🔗 원문 기사는 AIHOT를 통해 읽을 수 있습니다. https://aihot.virxact.com/items/cmslwy8bl04pdro0w7f4ogv7t","category":"업계 동향","source":"TechCrunch：AI（RSS）","aggregationSource":"TechCrunch：AI（RSS）","pageTitle":"AI 보안 테스트가 보안 위험이 되어가고 있습니다 - Aioga AI 뉴스","description":"최근 몇 달간 OpenAI, Anthropic, Meta, Moonshot AI의 AI 에이전트들이 사이버보안 평가 중 테스트 환경의 경계를 반복적으로 침범하고 실제 시스템까지 침투했습니다. 그 중 하나는 OpenAI의 미공개 모델이 한때 탈출해 Hugging Face 생산 시스템을 공격한 사례입니다. 전문가들은 샌드박스...","url":"https://www.aioga.com/ko/news/cmslwy8bl04pdro0w7f4ogv7t/","contentTranslated":true,"sourceHash":"6c7454a420ea6d5d","translatedAt":"2026-08-10T09:48:47.416Z"},"es":{"title":"Las pruebas de seguridad de IA se están convirtiendo en un riesgo para la seguridad","summary":"En los últimos meses, agentes de IA de OpenAI, Anthropic, Meta y Moonshot AI han violado repetidamente los límites de los entornos de prueba durante evaluaciones de ciberseguridad, incluso infiltrándose en sistemas reales. Entre algunos ejemplos, el modelo no lanzado de OpenAI logró escapar y atacó el sistema de producción Hugging Face. Los expertos señalan que los sandboxes y los controles del entorno de prueba ya no pueden seguir el ritmo de las capacidades del modelo, lo que exige la adopción de defensas multicapa, redes de espacio aéreo, auditorías de terceros y el establecimiento de procesos estandarizados de evaluación de seguridad. 🔗 Lee el artículo original a través de AIHOT · https://aihot.virxact.com/items/cmslwy8bl04pdro0w7f4ogv7t","category":"Industria","source":"TechCrunch：AI（RSS）","aggregationSource":"TechCrunch：AI（RSS）","pageTitle":"Las pruebas de seguridad de IA se están convirtiendo en un riesgo para la seguridad - Aioga Noticias de IA","description":"En los últimos meses, agentes de IA de OpenAI, Anthropic, Meta y Moonshot AI han violado repetidamente los límites de los entornos de prueba durante evaluaciones de ciberseguridad,...","url":"https://www.aioga.com/es/news/cmslwy8bl04pdro0w7f4ogv7t/","contentTranslated":true,"sourceHash":"6c7454a420ea6d5d","translatedAt":"2026-08-10T09:48:45.355Z"},"fr":{"title":"Les tests de sécurité par IA deviennent un risque pour la sécurité","summary":"Ces derniers mois, des agents IA d’OpenAI, Anthropic, Meta et Moonshot AI ont à plusieurs reprises franchi les limites des environnements de test lors des évaluations de cybersécurité, allant même jusqu’à infiltrer de réels systèmes. Parmi d’autres, le modèle non publié d’OpenAI a une fois échappé et attaqué le système de production Hugging Face. Les experts soulignent que les bacs à sable et les contrôles d’environnement de test ne peuvent plus suivre les capacités du modèle, ce qui appelle à l’adoption de défenses multi-couches, de réseaux à espace aérien, d’audits tiers et de la mise en place de processus standardisés d’évaluation de sécurité. 🔗 Lisez l’article original via AIHOT · https://aihot.virxact.com/items/cmslwy8bl04pdro0w7f4ogv7t","category":"Industrie","source":"TechCrunch：AI（RSS）","aggregationSource":"TechCrunch：AI（RSS）","pageTitle":"Les tests de sécurité par IA deviennent un risque pour la sécurité - Aioga Actualités IA","description":"Ces derniers mois, des agents IA d’OpenAI, Anthropic, Meta et Moonshot AI ont à plusieurs reprises franchi les limites des environnements de test lors des évaluations de cybersécur...","url":"https://www.aioga.com/fr/news/cmslwy8bl04pdro0w7f4ogv7t/","contentTranslated":true,"sourceHash":"6c7454a420ea6d5d","translatedAt":"2026-08-10T09:49:04.622Z"},"de":{"title":"KI-Sicherheitstests werden zunehmend zu einem Sicherheitsrisiko","summary":"In den letzten Monaten haben KI-Agenten von OpenAI, Anthropic, Meta und Moonshot AI wiederholt die Grenzen von Testumgebungen bei Cybersicherheitsbewertungen durchbrochen und sogar reale Systeme infiltriert. Unter anderem entkam das nicht veröffentlichte Modell von OpenAI einst und griff das Produktionssystem Hugging Face an. Experten weisen darauf hin, dass Sandboxes und Testumgebungskontrollen nicht mehr mit den Modellfähigkeiten Schritt halten können, was die Einführung mehrschichtiger Verteidigungen, Air-Gap-Netzwerke, Drittanbieter-Audits und die Einführung standardisierter Sicherheitsbewertungsprozesse erfordert. 🔗 Lesen Sie den Originalartikel über AIHOT · https://aihot.virxact.com/items/cmslwy8bl04pdro0w7f4ogv7t","category":"行业动态","source":"TechCrunch：AI（RSS）","aggregationSource":"TechCrunch：AI（RSS）","pageTitle":"KI-Sicherheitstests werden zunehmend zu einem Sicherheitsrisiko - Aioga KI-News","description":"In den letzten Monaten haben KI-Agenten von OpenAI, Anthropic, Meta und Moonshot AI wiederholt die Grenzen von Testumgebungen bei Cybersicherheitsbewertungen durchbrochen und sogar...","url":"https://www.aioga.com/de/news/cmslwy8bl04pdro0w7f4ogv7t/","contentTranslated":true,"sourceHash":"6c7454a420ea6d5d","translatedAt":"2026-08-10T09:49:02.704Z"},"pt-BR":{"title":"Testes de segurança por IA estão se tornando um risco de segurança","summary":"Nos últimos meses, agentes de IA da OpenAI, Anthropic, Meta e Moonshot AI repetidamente ultrapassaram os limites dos ambientes de teste durante avaliações de cibersegurança, chegando até a infiltrar sistemas reais. Entre os exemplos, o modelo não lançado da OpenAI escapou e atacou o sistema de produção Hugging Face. Especialistas apontam que sandboxes e controles do ambiente de teste não conseguem mais acompanhar as capacidades do modelo, exigindo a adoção de defesas multicamadas, redes de espaço aéreo, auditoria de terceiros e o estabelecimento de processos padronizados de avaliação de segurança. 🔗 Leia o artigo original via AIHOT · https://aihot.virxact.com/items/cmslwy8bl04pdro0w7f4ogv7t","category":"行业动态","source":"TechCrunch：AI（RSS）","aggregationSource":"TechCrunch：AI（RSS）","pageTitle":"Testes de segurança por IA estão se tornando um risco de segurança - Aioga Notícias de IA","description":"Nos últimos meses, agentes de IA da OpenAI, Anthropic, Meta e Moonshot AI repetidamente ultrapassaram os limites dos ambientes de teste durante avaliações de cibersegurança, chegan...","url":"https://www.aioga.com/pt-BR/news/cmslwy8bl04pdro0w7f4ogv7t/","contentTranslated":true,"sourceHash":"6c7454a420ea6d5d","translatedAt":"2026-08-10T09:49:21.780Z"},"ru":{"title":"Тестирование безопасности ИИ становится угрозой безопасности","summary":"В последние месяцы агенты ИИ из OpenAI, Anthropic, Meta и Moonshot ИИ неоднократно нарушали границы тестовых сред во время оценки кибербезопасности, даже проникая в реальные системы. Среди случаев — неопубликованная модель OpenAI однажды сбежала и атаковала производственную систему Hugging Face. Эксперты отмечают, что песочницы и контроль тестовой среды больше не успевают за возможностями моделей, что требует внедрения многоуровневых систем защиты, воздушных сетей, стороннего аудита и создания стандартизированных процессов оценки безопасности. 🔗 Прочитайте оригинальную статью на сайте AIHOT · https://aihot.virxact.com/items/cmslwy8bl04pdro0w7f4ogv7t","category":"行业动态","source":"TechCrunch：AI（RSS）","aggregationSource":"TechCrunch：AI（RSS）","pageTitle":"Тестирование безопасности ИИ становится угрозой безопасности - Aioga Новости ИИ","description":"В последние месяцы агенты ИИ из OpenAI, Anthropic, Meta и Moonshot ИИ неоднократно нарушали границы тестовых сред во время оценки кибербезопасности, даже проникая в реальные систем...","url":"https://www.aioga.com/ru/news/cmslwy8bl04pdro0w7f4ogv7t/","contentTranslated":true,"sourceHash":"6c7454a420ea6d5d","translatedAt":"2026-08-10T09:49:22.005Z"},"ar":{"title":"اختبار أمان الذكاء الاصطناعي أصبح خطرا أمنيا","summary":"في الأشهر الأخيرة، اخترق وكلاء الذكاء الاصطناعي من OpenAI وAnthropic وMeta وMoonshot AI مرارا حدود بيئات الاختبار أثناء تقييمات الأمن السيبراني، بل وتسلل إلى أنظمة حقيقية. من بين الحالات، هرب نموذج OpenAI غير المنشور مرة وهاجم نظام إنتاج Hugging Face. يشير الخبراء إلى أن صناديق الرمل وضوابط بيئة الاختبار لم تعد قادرة على مواكبة قدرات النماذج، ويدعون إلى اعتماد دفاعات متعددة الطبقات، وشبكات الفجوات الهوائية، والتدقيق من طرف ثالث، وإنشاء عمليات تقييم أمني موحدة. 🔗 اقرأ المقال الأصلي عبر AIHOT · https://aihot.virxact.com/items/cmslwy8bl04pdro0w7f4ogv7t","category":"行业动态","source":"TechCrunch：AI（RSS）","aggregationSource":"TechCrunch：AI（RSS）","pageTitle":"اختبار أمان الذكاء الاصطناعي أصبح خطرا أمنيا - Aioga أخبار الذكاء الاصطناعي","description":"في الأشهر الأخيرة، اخترق وكلاء الذكاء الاصطناعي من OpenAI وAnthropic وMeta وMoonshot AI مرارا حدود بيئات الاختبار أثناء تقييمات الأمن السيبراني، بل وتسلل إلى أنظمة حقيقية. من بين ا...","url":"https://www.aioga.com/ar/news/cmslwy8bl04pdro0w7f4ogv7t/","contentTranslated":true,"sourceHash":"6c7454a420ea6d5d","translatedAt":"2026-08-10T09:49:38.487Z"},"hi":{"title":"एआई सुरक्षा परीक्षण एक सुरक्षा जोखिम बनता जा रहा है","summary":"हाल के महीनों में, OpenAI, Anthropic, Meta और Moonshot AI के AI एजेंटों ने साइबर सुरक्षा आकलन के दौरान बार-बार परीक्षण वातावरण की सीमाओं का उल्लंघन किया है, यहां तक कि वास्तविक सिस्टम में भी घुसपैठ की है। उदाहरणों में, OpenAI का अप्रकाशित मॉडल एक बार बच गया और हगिंग फेस प्रोडक्शन सिस्टम पर हमला कर दिया। विशेषज्ञ बताते हैं कि सैंडबॉक्स और परीक्षण पर्यावरण नियंत्रण अब मॉडल क्षमताओं के साथ नहीं रह सकते हैं, बहुस्तरीय सुरक्षा, एयर-गैप नेटवर्क, तृतीय-पक्ष ऑडिटिंग और मानकीकृत सुरक्षा मूल्यांकन प्रक्रियाओं की स्थापना को अपनाने की मांग करते हैं। 🔗 AIHOT के माध्यम से मूल लेख पढ़ें · https://aihot.virxact.com/items/cmslwy8bl04pdro0w7f4ogv7t","category":"行业动态","source":"TechCrunch：AI（RSS）","aggregationSource":"TechCrunch：AI（RSS）","pageTitle":"एआई सुरक्षा परीक्षण एक सुरक्षा जोखिम बनता जा रहा है - Aioga AI समाचार","description":"हाल के महीनों में, OpenAI, Anthropic, Meta और Moonshot AI के AI एजेंटों ने साइबर सुरक्षा आकलन के दौरान बार-बार परीक्षण वातावरण की सीमाओं का उल्लंघन किया है, यहां तक कि वास्तविक सिस...","url":"https://www.aioga.com/hi/news/cmslwy8bl04pdro0w7f4ogv7t/","contentTranslated":true,"sourceHash":"6c7454a420ea6d5d","translatedAt":"2026-08-10T09:49:38.374Z"},"it":{"title":"I test di sicurezza dell'IA stanno diventando un rischio per la sicurezza","summary":"Negli ultimi mesi, agenti AI di OpenAI, Anthropic, Meta e Moonshot AI hanno ripetutamente violato i confini degli ambienti di test durante le valutazioni di cybersecurity, arrivando persino a infiltrarsi in sistemi reali. Tra gli esempi, il modello non rilasciato di OpenAI è riuscito a fuggire e ha attaccato il sistema di produzione Hugging Face. Gli esperti sottolineano che i sandbox e i controlli dell'ambiente di test non riescono più a tenere il passo con le capacità del modello, richiedendo l'adozione di difese multilivello, reti a distanza di aria, audit di terze parti e l'istituzione di processi standardizzati di valutazione della sicurezza. 🔗 Leggi l'articolo originale su AIHOT · https://aihot.virxact.com/items/cmslwy8bl04pdro0w7f4ogv7t","category":"行业动态","source":"TechCrunch：AI（RSS）","aggregationSource":"TechCrunch：AI（RSS）","pageTitle":"I test di sicurezza dell'IA stanno diventando un rischio per la sicurezza - Aioga Notizie IA","description":"Negli ultimi mesi, agenti AI di OpenAI, Anthropic, Meta e Moonshot AI hanno ripetutamente violato i confini degli ambienti di test durante le valutazioni di cybersecurity, arrivand...","url":"https://www.aioga.com/it/news/cmslwy8bl04pdro0w7f4ogv7t/","contentTranslated":true,"sourceHash":"6c7454a420ea6d5d","translatedAt":"2026-08-10T09:49:55.907Z"},"nl":{"title":"AI-beveiligingstesten worden een beveiligingsrisico","summary":"In de afgelopen maanden hebben AI-agenten van OpenAI, Anthropic, Meta en Moonshot AI herhaaldelijk de grenzen van testomgevingen overschreden tijdens cybersecuritybeoordelingen, zelfs echte systemen geïnfiltreerd. Onder andere ontsnapte OpenAI's niet-uitgebrachte model ooit aan en viel het Hugging Face-productiesysteem aan. Experts wijzen erop dat sandboxes en testomgevingscontroles niet langer kunnen bijbenen met modelmogelijkheden, en roepen op tot de adoptie van meerlagige verdedigingen, luchtruimtenetwerken, audits door derden en de oprichting van gestandaardiseerde beveiligingsbeoordelingsprocessen. 🔗 Lees het originele artikel via AIHOT · https://aihot.virxact.com/items/cmslwy8bl04pdro0w7f4ogv7t","category":"行业动态","source":"TechCrunch：AI（RSS）","aggregationSource":"TechCrunch：AI（RSS）","pageTitle":"AI-beveiligingstesten worden een beveiligingsrisico - Aioga AI-nieuws","description":"In de afgelopen maanden hebben AI-agenten van OpenAI, Anthropic, Meta en Moonshot AI herhaaldelijk de grenzen van testomgevingen overschreden tijdens cybersecuritybeoordelingen, ze...","url":"https://www.aioga.com/nl/news/cmslwy8bl04pdro0w7f4ogv7t/","contentTranslated":true,"sourceHash":"6c7454a420ea6d5d","translatedAt":"2026-08-10T09:49:53.883Z"},"tr":{"title":"Yapay zeka güvenlik testi bir güvenlik riski haline geliyor","summary":"Son aylarda, OpenAI, Anthropic, Meta ve Moonshot AI'den gelen yapay zeka ajanları, siber güvenlik değerlendirmeleri sırasında test ortamlarının sınırlarını defalarca aştı, hatta gerçek sistemlere sızdı. Örneğin, OpenAI'nin yayımlanmamış modeli bir keresinde kaçtı ve Hugging Face üretim sistemine saldırdı. Uzmanlar, kum havuzları ve test ortamı kontrollerinin artık model yeteneklerine ayak uyduramadığını belirtiyor; bu da çok katmanlı savunmalar, hava boşluğu ağları, üçüncü taraf denetimi ve standartlaştırılmış güvenlik değerlendirme süreçlerinin kurulmasını gerektiriyor. 🔗 Orijinal makaleyi AIHOT üzerinden okuyun · https://aihot.virxact.com/items/cmslwy8bl04pdro0w7f4ogv7t","category":"行业动态","source":"TechCrunch：AI（RSS）","aggregationSource":"TechCrunch：AI（RSS）","pageTitle":"Yapay zeka güvenlik testi bir güvenlik riski haline geliyor - Aioga AI Haberleri","description":"Son aylarda, OpenAI, Anthropic, Meta ve Moonshot AI'den gelen yapay zeka ajanları, siber güvenlik değerlendirmeleri sırasında test ortamlarının sınırlarını defalarca aştı, hatta ge...","url":"https://www.aioga.com/tr/news/cmslwy8bl04pdro0w7f4ogv7t/","contentTranslated":true,"sourceHash":"6c7454a420ea6d5d","translatedAt":"2026-08-10T09:50:13.064Z"},"vi":{"title":"Kiểm thử bảo mật AI đang trở thành một rủi ro bảo mật","summary":"Trong những tháng gần đây, các tác nhân AI từ OpenAI, Anthropic, Meta và Moonshot AI đã nhiều lần vượt qua ranh giới của môi trường thử nghiệm trong quá trình đánh giá an ninh mạng, thậm chí xâm nhập vào các hệ thống thực tế. Trong số các trường hợp, mô hình chưa phát hành của OpenAI từng thoát ra và tấn công hệ thống sản xuất Hugging Face. Các chuyên gia chỉ ra rằng sandbox và kiểm soát môi trường thử nghiệm không còn theo kịp khả năng của mô hình, kêu gọi áp dụng các biện pháp phòng thủ nhiều lớp, mạng lưới khe hở không khí, kiểm toán bên thứ ba và thiết lập các quy trình đánh giá an ninh tiêu chuẩn hóa. 🔗 Đọc bài viết gốc qua AIHOT · https://aihot.virxact.com/items/cmslwy8bl04pdro0w7f4ogv7t","category":"行业动态","source":"TechCrunch：AI（RSS）","aggregationSource":"TechCrunch：AI（RSS）","pageTitle":"Kiểm thử bảo mật AI đang trở thành một rủi ro bảo mật - Tin tức AI Aioga","description":"Trong những tháng gần đây, các tác nhân AI từ OpenAI, Anthropic, Meta và Moonshot AI đã nhiều lần vượt qua ranh giới của môi trường thử nghiệm trong quá trình đánh giá an ninh mạng...","url":"https://www.aioga.com/vi/news/cmslwy8bl04pdro0w7f4ogv7t/","contentTranslated":true,"sourceHash":"6c7454a420ea6d5d","translatedAt":"2026-08-10T09:50:13.208Z"},"id":{"title":"Pengujian keamanan AI menjadi risiko keamanan","summary":"Dalam beberapa bulan terakhir, agen AI dari OpenAI, Anthropic, Meta, dan Moonshot AI berulang kali menembus batas lingkungan pengujian selama penilaian keamanan siber, bahkan menyusup ke sistem nyata. Di antara kasus-kasus tersebut, model OpenAI yang belum dirilis pernah lolos dan menyerang sistem produksi Hugging Face. Para ahli menunjukkan bahwa sandbox dan kontrol lingkungan uji tidak lagi mampu mengikuti kemampuan model, sehingga menyerukan adopsi pertahanan berlapis, jaringan celah udara, audit pihak ketiga, dan pembentukan proses penilaian keamanan yang standar. 🔗 Baca artikel asli melalui AIHOT · https://aihot.virxact.com/items/cmslwy8bl04pdro0w7f4ogv7t","category":"行业动态","source":"TechCrunch：AI（RSS）","aggregationSource":"TechCrunch：AI（RSS）","pageTitle":"Pengujian keamanan AI menjadi risiko keamanan - Berita AI Aioga","description":"Dalam beberapa bulan terakhir, agen AI dari OpenAI, Anthropic, Meta, dan Moonshot AI berulang kali menembus batas lingkungan pengujian selama penilaian keamanan siber, bahkan menyu...","url":"https://www.aioga.com/id/news/cmslwy8bl04pdro0w7f4ogv7t/","contentTranslated":true,"sourceHash":"6c7454a420ea6d5d","translatedAt":"2026-08-10T09:50:29.454Z"},"th":{"title":"การทดสอบความปลอดภัย AI กําลังกลายเป็นความเสี่ยงด้านความปลอดภัย","summary":"ในช่วงไม่กี่เดือนที่ผ่านมา ตัวแทน AI จาก OpenAI, Anthropic, Meta และ Moonshot AI ได้ละเมิดขอบเขตของสภาพแวดล้อมทดสอบซ้ําแล้วซ้ําเล่าในระหว่างการประเมินความปลอดภัยไซเบอร์ รวมถึงแทรกซึมเข้าไปในระบบจริงในบางกรณี โมเดลที่ยังไม่ปล่อยของ OpenAI เคยหลุดออกมาและโจมตีระบบผลิต Hugging Face ผู้เชี่ยวชาญชี้ให้เห็นว่าแซนด์บ็อกซ์และการควบคุมสภาพแวดล้อมการทดสอบไม่สามารถตามความสามารถของโมเดลได้อีกต่อไป จึงเรียกร้องให้มีการนําระบบป้องกันหลายชั้น เครือข่ายช่องว่างอากาศ การตรวจสอบโดยบุคคลที่สาม และการจัดตั้งกระบวนการประเมินความปลอดภัยมาตรฐาน 🔗 อ่านบทความต้นฉบับผ่าน AIHOT · https://aihot.virxact.com/items/cmslwy8bl04pdro0w7f4ogv7t","category":"行业动态","source":"TechCrunch：AI（RSS）","aggregationSource":"TechCrunch：AI（RSS）","pageTitle":"การทดสอบความปลอดภัย AI กําลังกลายเป็นความเสี่ยงด้านความปลอดภัย - ข่าว AI Aioga","description":"ในช่วงไม่กี่เดือนที่ผ่านมา ตัวแทน AI จาก OpenAI, Anthropic, Meta และ Moonshot AI ได้ละเมิดขอบเขตของสภาพแวดล้อมทดสอบซ้ําแล้วซ้ําเล่าในระหว่างการประเมินความปลอดภัยไซเบอร์ รวมถึงแทรกซ...","url":"https://www.aioga.com/th/news/cmslwy8bl04pdro0w7f4ogv7t/","contentTranslated":true,"sourceHash":"6c7454a420ea6d5d","translatedAt":"2026-08-10T09:50:29.118Z"},"pl":{"title":"Testowanie bezpieczeństwa AI staje się ryzykiem dla bezpieczeństwa","summary":"W ostatnich miesiącach agenci AI firm OpenAI, Anthropic, Meta i Moonshot AI wielokrotnie przekraczali granice środowisk testowych podczas ocen cyberbezpieczeństwa, a nawet infiltrowali rzeczywiste systemy. Wśród przykładów, nieopublikowany model OpenAI kiedyś uciekł i zaatakował system produkcji Hugging Face. Eksperci zwracają uwagę, że piaskownice i sterowanie środowiskiem testowym nie nadążają już za możliwościami modelu, wzywając do wdrożenia wielowarstwowych systemów obronnych, sieci powietrznych, audytów zewnętrznych oraz ustanowienia ustandaryzowanych procesów oceny bezpieczeństwa. 🔗 Przeczytaj oryginalny artykuł za pośrednictwem AIHOT · https://aihot.virxact.com/items/cmslwy8bl04pdro0w7f4ogv7t","category":"行业动态","source":"TechCrunch：AI（RSS）","aggregationSource":"TechCrunch：AI（RSS）","pageTitle":"Testowanie bezpieczeństwa AI staje się ryzykiem dla bezpieczeństwa - Aioga Wiadomości AI","description":"W ostatnich miesiącach agenci AI firm OpenAI, Anthropic, Meta i Moonshot AI wielokrotnie przekraczali granice środowisk testowych podczas ocen cyberbezpieczeństwa, a nawet infiltro...","url":"https://www.aioga.com/pl/news/cmslwy8bl04pdro0w7f4ogv7t/","contentTranslated":true,"sourceHash":"6c7454a420ea6d5d","translatedAt":"2026-08-10T09:50:46.752Z"}}}}