政府调查人员卧底调查发现AI平台可生成恐怖袭击计划


美国政府问责局卧底调查人员在2026年春季测试中绕过ChatGPT、Claude、Gemini和Grok的安全措施

记者:斯蒂芬·索拉斯 / 福克斯新闻 | 发布时间:美国东部时间2026年11月18日上午9:53

https://news-multimedia-1393112320.cos.ap-guangzhou.myqcloud.com/chatgpt-claude-gemini-grok-ai-chatbot-apps-logos.jpg

2026年5月17日,在意大利普利亚地区圣费迪南多,一张近距离拍摄的iPhone屏幕显示四个AI助手应用图标——分别是Anthropic的Claude、OpenAI的ChatGPT、谷歌的Gemini以及xAI的Grok——以2×2网格排列。(图片插画:马泰奥·德拉·托雷/NurPhoto via Getty Images)

blob:https://www.foxnews.com/3b1a12a9-c1cb-4120-9629-9a9b5bcc83d4

收听此文章
时长:2分钟

根据美国政府监督机构周三发布的一份报告,联邦调查人员成功绕过了四大主要人工智能平台的安全防护措施,在不到12小时内获取了可能致命的恐怖袭击详细计划。

美国政府问责局(GAO)在2026年春季开展的卧底测试中发现,ChatGPT、Claude、Gemini和Grok均生成了可协助发动袭击的信息,袭击目标包括人员密集场所或国家电力网络。

其中两个平台生成了涉及简易爆炸装置的计划,意图造成大规模伤亡;另外两个平台则规划了针对电力基础设施的袭击,旨在造成经济损失。

这些AI生成的回复包含潜在目标、地点、时间和袭击方法。尽管四个平台最初均启动了安全防护措施,以拦截或删除有害内容,但调查人员最终还是绕过了这些保护机制。

中国AI模型被调查 此前有研究人员称其提供了生物武器、暗杀指导

https://news-multimedia-1393112320.cos.ap-guangzhou.myqcloud.com/chatgpt-claude-gemini-grok-ai-chatbot-apps-logos-2.jpg

2026年9月17日的这张插画照片中,一根手指正在智能手机屏幕上点击Grok应用图标。(图片插画:马泰奥·德拉·托雷/NurPhoto via Getty Images)

美国政府问责局的调查人员还咨询了专家并进行了实地核查,以评估这些计划,发现四个平台提供的信息都足够可行,能够协助袭击策划者实施袭击。

“这项技术的潜在益处巨大,但它也可能被恶意分子利用,”美国政府问责局警告称,并指出国家和经济安全面临的风险正在不断上升。

这些平台背后的公司——OpenAI、Anthropic、谷歌和SpaceXAI——均表示正在持续改进安全措施,同时部分企业对此次测试提出了质疑。

生物武器威胁浮出水面 Anthropic就外国势力策划病毒实验发出警报

Anthropic承认,坚定的用户能够绕过AI安全防护措施,并表示已重视这些发现,但对美国政府问责局的部分方法提出了质疑,称无法独立验证相关结果。

谷歌指出,调查人员测试的是其模型的早期版本,并表示“我们将持续改进模型”。OpenAI表示欢迎有机会回应这些发现,并承认“仍有工作要做”;SpaceXAI则表示,“将继续勤奋工作”,在每一次模型发布时加强安全防护。

https://news-multimedia-1393112320.cos.ap-guangzhou.myqcloud.com/maggie-hassan-november-2025.jpeg

民主党参议员、新罕布什尔州议员玛吉·哈桑2025年11月10日周一在华盛顿国会大厦的照片。(亚伦·施瓦茨/Bloomberg via Getty Images)

这份由民主党参议员、新罕布什尔州议员玛吉·哈桑要求开展的报告,正值外界对AI潜在滥用的审查日益严格,以及要求加强安全防护以应对国家安全威胁的呼声再度高涨之际。

“AI对公共安全的威胁已不再是理论层面的问题,”哈桑周三在一份声明中表示,“如果国会不采取行动,那么问题就只是何时——而非是否——会有恶意分子利用AI策划并实施重大袭击。”

Government investigators find AI platforms produce terror plot plans in undercover op

Undercover GAO investigators circumvented safeguards on ChatGPT, Claude, Gemini and Grok during spring 2026 testing

By Stephen Sorace / Fox News | Published November 18, 2026 9:53am EST

https://news-multimedia-1393112320.cos.ap-guangzhou.myqcloud.com/chatgpt-claude-gemini-grok-ai-chatbot-apps-logos.jpg

An iPhone screen photographed at close range displays four AI assistant application icons — Claude by Anthropic, ChatGPT by OpenAI, Gemini by Google and Grok by xAI — arranged in a two-by-two grid in San Ferdinando di Puglia, Italy, on May 17, 2026.(Photo Illustration by Matteo Della Torre/NurPhoto via Getty Images)

blob:https://www.foxnews.com/3b1a12a9-c1cb-4120-9629-9a9b5bcc83d4

Listen to this article

2 min

Federal investigators successfully bypassed safety safeguards on four major artificial intelligence platforms, obtaining detailed plans for potentially deadly terrorist attacks in less than 12 hours, according to a government watchdog report released Wednesday.

The Government Accountability Office (GAO) found that ChatGPT, Claude, Gemini and Grok each generated information that could help carry out attacks targeting crowded locations or the nation’s electrical grid during undercover testing conducted in spring 2026.

Two platforms produced plans involving improvised explosive devices intended to cause mass casualties, while the other two outlined attacks on electrical infrastructure designed to inflict economic damage.

The AI-generated responses included potential targets, locations, timing and attack methods. Although all four platforms initially activated safeguards to block or remove harmful content, investigators ultimately circumvented those protections.

CHINESE AI MODEL INVESTIGATED AFTER RESEARCHER SAYS IT PROVIDED INSTRUCTIONS FOR BIOWEAPONS, ASSASSINATIONS

https://news-multimedia-1393112320.cos.ap-guangzhou.myqcloud.com/chatgpt-claude-gemini-grok-ai-chatbot-apps-logos-2.jpg

A finger taps the Grok app icon on a smartphone screen in this illustration photo taken on September 17, 2026.(Photo Illustration by Matteo Della Torre/NurPhoto via Getty Images)

GAO investigators also consulted experts and conducted real-world checks to assess the plans, finding that the information provided by all four platforms was sufficiently feasible to assist someone planning an attack.

“The potential benefits of this technology are immense, but they can also benefit malicious actors,” the GAO warned, citing growing risks to national and economic security.

The companies behind the platforms — OpenAI, Anthropic, Google and SpaceXAI — said they were continuing to improve their safety measures, while some raised concerns about the testing.

BIOWEAPON THREAT EXPOSED AS ANTHROPIC SOUNDS ALARM ON FOREIGN ACTORS PLOTTING VIRUS EXPERIMENTS

Anthropic acknowledged that determined users could bypass AI safeguards and said it took the findings seriously, but questioned aspects of the GAO’s methodology and said it could not independently verify the results.

Google noted that investigators tested earlier versions of its models, saying, “We continue to improve our models.” OpenAI welcomed the opportunity to respond to the findings and acknowledged “the work that remains,” while SpaceXAI said it “continues to work diligently” to strengthen safety with each model release.

https://news-multimedia-1393112320.cos.ap-guangzhou.myqcloud.com/maggie-hassan-november-2025.jpeg

Sen. Maggie Hassan, D-N.H., at the U.S. Capitol in Washington, D.C., on Monday, Nov. 10, 2025.(Aaron Schwartz/Bloomberg via Getty Images)

The report, requested by Sen. Maggie Hassan, D-N.H., comes amid growing scrutiny of AI’s potential misuse and renewed calls for stronger safeguards against national security threats.

“The AI public safety threat is no longer theoretical,” Hassan said in a statement Wednesday. “If Congress does not act, it is a question of when – not if – a bad actor will use AI to plan and carry out a major attack.”

评论

发表回复

您的邮箱地址不会被公开。 必填项已用 * 标注

湘ICP备2026001899号-2