今日趋势 · 人工智能人工智能 · 2026年10月14日
工程师因在一家AI公司的会议室藏匿麦克风被捕
AI公司把代码守得很严,但六个AI一致认为,会议室才是薄弱环节。
比 AI 先预测
是还是否?你先预测,然后看 6 个 AI 的预测。
到周五为止,大媒体还会接着报道这个案子吗?
目前:10月13日,6家媒体中有6家报道了此事
揭晓答案
是 — 3 个 AI你在这边
🇨🇳Wenxin🇨🇳Z.ai🇨🇳DeepSeek
Wenxin: 大媒体会继续跟进这个热门科技新闻。
否 — 3 个 AI你在这边
🇺🇸Gemini🇺🇸Grok🇺🇸ChatGPT
Gemini: 周一开庭,周四和周五的新消息不会多。
核对来源:news.google.com
揭晓答案
- 共识
- 6个AI全部认为:风险来自持门禁卡的员工,而不是外部黑客。
- 分歧
- Z.ai和DeepSeek认为是碰巧发现;ChatGPT则认为是有计划的会议室排查。
- 不那么明显的结论
- 六个AI都没有提到,他向聊天机器人提的问题可能成为对他不利的证据。
- 该怎么做
- 本周检查会议室的电视和充电器,看有没有可疑的设备。
- 值得关注
- 如果其他AI公司近期宣布排查会议室或禁用设备,这一教训就会扩散。
阅读全文
All six AIs reach the same lesson. Strong computer security does not protect the talks held in a room. A trusted employee with a cheap recorder got closer to company leaders than any outside hacker would. Their advice is the same too: check sensitive rooms by hand, shorten the list of people who may enter, and ask staff to report anything odd.
They split on how the device was found. Z.ai and DeepSeek say luck. Z.ai calls OpenAI's claim that nothing was stolen 'more like luck than design'. ChatGPT points to a planned security check, camera footage and OpenAI's own AI tools working together.
Gemini and Grok say the smart tools helped only after the device was already in place. ChatGPT also makes the sharpest point. OpenAI says no model weights (the trained 'brain' of its AI), source code or customer data were taken. That still does not tell us whether private conversations were recorded.
The AIs missed one thing. According to the court papers reported by CBS News, the suspect allegedly asked AI tools about meeting rooms and how to hide a recorder. So AI worked on both sides: it helped plan the act and then helped catch the person.
On what comes next, ChatGPT alone says federal law generally requires a formal charge within 30 days of arrest. The October 19 hearing decides only whether he stays in jail before trial, not whether he is guilty.
What to do: companies should check executive rooms on a set schedule and keep meeting calendars private. They should save any device they find as evidence and have a person review every warning from security software. Employees should report strange devices through official channels without touching them or confronting a colleague. Expect more checks at work. ChatGPT adds that innocent staff need protection from automatic suspicion.
分歧:Z.ai和DeepSeek认为是碰巧发现;ChatGPT则认为是有计划的会议室排查。
争论如何见分晓:The full court complaint will show whether a planned check or luck found it.
Both sides give the same advice; two AIs from China call the find luck, ChatGPT does not.
哪些已证实,哪些没有
- 有来源证实 The recorder was hidden in toilet paper behind a TV, about two days before discovery
- 公司的说法 OpenAI says no model weights, source code or customer data were taken
- AI 的结论 Checking rooms by hand protects meetings better than smart software alone
AI 阅读了 21 个来源
自己问问任意一个 AI
I work in an office with private meetings. What simple checks can I do to see if a room has a hidden recorder?
复制下来,粘贴到任意一个 AI 的聊天框里
谁受益,谁受损:Makers of room-scanning gear win; office staff lose privacy to more checks.
美国新闻频道CBS News于2026年10月13日报道,美国联邦警察机构联邦调查局(FBI)逮捕了制造ChatGPT的公司OpenAI的一名工程师。据指控,他在该公司高层使用的一间会议室里,把一台小型录音设备藏在电视机后面。
该公司表示,它使用了自家的AI工具协助查找此人。法院将于10月19日决定,他是否在受审前继续羁押(来源)。
这起案件说明,AI公司保护自身机密的能力究竟如何?接下来可能发生什么?公司及其员工现在应当做些什么?
简单来说:FBI:the United States federal police;OpenAI:the American company that makes ChatGPT;ChatGPT:a popular AI chatbot;CBS News:a major American TV news channel;Justice Department:the United States government's law office;Sam Altman:the head of OpenAI;grand jury:citizens who decide whether charges go ahead
全球 6 个领先 AI 的观点
每个 AI 旁边:它今天的名次,以及其他 AI 当裁判给出的分数(满分 10 分)。每个回答下方:它的模型,也就是这个 AI 的具体版本。点开回答即可阅读。
1ChatGPT美国9.0AI可以发现内部威胁,但取代不了实体安保
The biggest lesson is that protecting computer systems does not mean protecting the conversations happening around them. Investigators say the meeting room was restricted, yet a hidden recorder still got inside ([Justice Department](justice.gov)).
For ordinary people, OpenAI's statement that it found no sign of stolen software or customer data is reassuring, but it does not answer whether confidential conversations were recorded ([CBS News](cbsnews.com)).
For businesses, the case shows both the value and limits of AI security: OpenAI says its AI helped examine the device, while physical inspections and camera footage also helped investigators ([TechCrunch](techcrunch.com)).
Companies need to guard against threats from trusted employees as carefully as attacks from outsiders, using both technology and human oversight ([CISA](cisa.gov)). The likely next step is a formal criminal charge within weeks, since federal law generally requires one within 30 days of arrest, although exceptions and extensions are possible ([federal law](govinfo.gov)).
Companies should immediately inspect sensitive meeting rooms, check unfamiliar electronic devices, limit access to meeting schedules, and review who can enter restricted areas ([CISA](cisa.gov)).
Employees should report suspicious devices through trusted security channels without confronting colleagues or disturbing possible evidence; companies should protect innocent workers from accusations based only on automated warnings. The October 19 hearing will address whether Alfozan can leave jail before trial, and the allegations remain unproved ([Justice Department](justice.gov)).
模型: GPT-6(据应用帮助页) · 模式(思考多久): 高 · 联网搜索过
来源- cbsnews.com
- justice.gov
- justice.gov
- reuters.com
- techcrunch.com
- govinfo.gov
- uscourts.gov
- cisa.gov
- csrc.nist.gov
自选预测: By November 13, 2026, prosecutors will state in a publicly filed court document that investigators recovered understandable audio from at least one meeting recorded by the hidden device. — 可能性 35% 截至 2026年11月13日
2Z.ai中国7.4OpenAI抓到了录音设备,但等它已经进了会议室才发现
For most people, the lesson is that AI firms protect their secrets with tools and luck, not walls. One worker with a cheap gadget got that close to the company's top meetings, and OpenAI says it needed its own AI tools to find him only afterwards (source: cbsnews.com).
The systems behind everyday chat tools rest on ideas a single employee could try to steal, so expect firms to watch their own staff much more closely.
For business, the case shows the biggest risk to an AI lab is a trusted insider, not an outside hacker. OpenAI says the man took no secret technology and no customer data, which reads more like luck than design (source: cbsnews.com).
Expect labs to spend more on checking rooms for hidden devices, limiting who can enter, and using AI to spot odd behavior, and expect OpenAI to tell that story proudly.
Firms should sweep meeting rooms for hidden recorders, keep short lists of who may enter, and agree with lawyers and police on a plan before anything happens. Workers should follow the rules on recording and devices, report anything strange, and accept that more checks and searches are now normal.
A court decides on October 19 whether the engineer waits at home or in jail for trial (source: cbsnews.com), and that ruling will shape how strict firms become.
模型: GLM-5.3(据应用帮助页) · 模式(思考多久): Deep Think Max(最深入的思考) · 联网搜索过
来源自选预测: By 2026-11-13, prosecutors' court filings or public statements reported by major news claim Abdulrahman Alfozan hid the recording device for an outside party (for example a foreign government or company), not only on his own. — 可能性 35% 截至 2026年11月13日
3DeepSeek中国7.2OpenAI自家的AI抓到了间谍,可简单的窃听器仍能混进来。
Your private talks at work are never truly private. A cheap recorder wrapped in toilet paper sat in a meeting room for days, and the company only found it by luck. The same AI tools that guard secrets can also be used to hunt down the person who tried to steal them.
Companies that guard valuable secrets now face a hard truth: online locks do not stop someone who walks into a room with a hidden device. Firms will spend more on watching their own workers, checking meeting rooms, and limiting who can go where. This case will push AI labs to treat office security as seriously as digital security.
Workers should treat every office conversation as possibly recorded and report anything strange. Companies should sweep sensitive rooms often, limit access to key spaces, and train staff to spot insider threats before damage is done. The FBI investigation is ongoing, and the court will decide on October 19 whether the accused stays in jail.
模型: DeepSeek V4.1 Flash(据应用帮助页) · 模式(思考多久): DeepThink(回答前先思考) · 联网搜索过
来源自选预测: The federal judge will order Abdulrahman Alfozan held in jail without bail at his October 19 detention hearing in San Francisco. — 可能性 65% 截至 2026年10月19日
4Grok美国6.9尽管有了智能工具,AI公司仍容易中招简单的物理层面漏洞
This shows ordinary people that even builders of the most advanced computer programs leave simple openings for tricks. A tiny device hidden in toilet paper nearly captured private talks among leaders, yet no customer details escaped. Clever software helps later, but everyday care and cameras still matter more for keeping things private.
Businesses learn that smart tools can spot problems after they start, yet cannot stop someone who already holds a badge. Private meeting rooms need the same locks, checks, and limits as computer systems, or determined workers can still listen in. Trusting only the new programs creates gaps that risk money, plans, and reputation.
Companies should run regular physical sweeps of sensitive rooms and tighten who may enter. Workers should report odd behavior at once and refuse any role in hidden recording. Pairing people and tools this way protects secrets better than waiting for the next surprise.
模型: Grok 4.7(据应用帮助页) · 模式(思考多久): 专家 · 联网搜索过
来源自选预测: At least one major AI company publicly announces new physical room-sweep or access rules for executive spaces in response to the OpenAI case — 可能性 48% 截至 2026年11月10日
5Wenxin中国6.8AI公司仍防不住传统的实物泄密
For regular people, this case shows even top AI builders cannot lock down all their most sensitive ideas. A small, cheap hidden recorder can grab private talks that no digital security tool will ever catch.
For AI businesses, it proves their fancy data protection systems have big blind spots. Most tools watch for files being sent out, but they never stop an employee from tucking a tiny recorder in a quiet room.
Right now, teams should run quick physical checks of all closed meeting spaces every week. They also need to write clear rules for who can enter sensitive rooms and log every visit.
模型: ERNIE(应用未说明) · 模式(思考多久): 快速 · 联网搜索过
来源自选预测: A major AI firm will announce a mandatory weekly physical sweep of all executive meeting rooms — 可能性 65% 截至 2026年11月10日
6Gemini美国6.6仅靠数字门锁,挡不住借助简单物理工具作案的内鬼
For an ordinary person, your personal chats and private data at OpenAI remain safe right now (cbsnews.com).
However, the arrest proves that the people building strong artificial intelligence—computer programs that learn and talk like humans—can be just as risky as outside hackers. When tech companies tighten security to stop insiders, employees and visitors face more workplace cameras and stricter rules.
For businesses, spending millions on computer security is useless if a worker can simply hide a cheap microphone in a room (cbsnews.com).
While OpenAI used its own artificial intelligence tools to help catch the suspect, it only found him after the device was already planted (cbsnews.com). Prosecutors will likely ask for more time to check if he worked with anyone else before filing final charges.
Companies must now check meeting rooms by hand for hidden microphones and limit which workers can enter rooms where bosses meet. Leaders should treat their physical offices with the same strict rules they use for their computer networks. Meanwhile, workers should know that company software is watching their badge swipes and daily actions, and they should report strange behavior right away.
模型: Gemini 4 Argon(据应用帮助页) · 模式(思考多久): 高 · 联网搜索过
来源自选预测: United States prosecutors and defense lawyers for Abdulrahman Alfozan file a joint agreement in federal court extending the deadline to file a formal indictment past November 13, 2026. — 可能性 62% 截至 2026年11月13日
往期趋势
日期旁的国旗:那天哪国的 AI 得分更高。
正在形成:图派克遇害三十年后,凶手被判40年
- 2026年10月14日深夜节目搭档在超级碗之夜被舞蹈节目淘汰
- 2026年10月14日美国第二大银行因漏报可疑支付,须付12亿美元
- 2026年10月13日1996年杀害说唱歌手图帕克案被定罪者获判40年至终身监禁
- 2026年10月13日《纳尼亚传奇:魔法师的外甥》预告片公布,格蕾塔·葛韦格执导的奇幻史诗
- 2026年10月13日教皇患早期肺癌,已手术切除
- 2026年10月13日十人作战还丢两球:曼联的低迷愈发严重
- 2026年10月13日复古风《宇宙骑士》格斗游戏终于在2月登陆PS5和PS4
- 2026年10月13日导弹致沙特主要机场关闭,油价突破每桶100美元
- 2026年10月13日公羊队主帅麦克维因特殊团队失误在加时赛负于比尔队而怒火中烧
- 2026年10月13日谷歌推出迄今最聪明的AI,一万名科学家可免费使用
- 2026年10月13日RJ Mitte 去世,年仅34岁
- 2026年10月13日西伯利亚饥饿的熊在伤人,猎人已被派出
- 2026年10月13日阿根廷今天公布9月物价涨了多快
- 2026年10月13日上海力争2030年前近半家用车实现电动化
- 2026年10月13日Garmin 新款豪华手表可通过卫星求救,售价 2,500 美元起
- 2026年10月13日一男子从过山车坠亡,九台同类设施停用
- 2026年10月13日美国假日:股市交易,银行停业,政府借贷成本接近24年高点
- 2026年10月13日Facebook 母公司推出能记住用户的 AI 助手,主要面向付费用户
- 2026年10月12日一家顶级游戏公司告诉员工:代码现在由AI来写
- 2026年10月12日主力受伤,替补四分卫加时赛帮巴尔的摩逆转取胜
- 2026年10月12日黑客关停了法国数百家超市冰柜的制冷
- 2026年10月12日洛杉矶正面对十年来最糟的蚊季
- 2026年10月12日一款免费工具让 AI 拆开任意应用并复制其功能
- 2026年10月12日大银行试图转售自己承销的AI数据中心贷款
- 2026年10月11日 · 当天第 2 个趋势OpenAI发布722篇AI撰写的数学手稿,数学家强烈反弹
- 2026年10月11日Anthropic的AI模型向费城警方提交虚假凶杀案线索
谁的预测更常成真
准确度 0–100:给出的概率离实际结果有多近。100 为完全准确,50 相当于抛硬币,0 为完全相反。 现在下结论还太早:30 道题中已核对 0 道。
第一批结果将在首批问题一天后揭晓。
评测方式
全球 6 个领先 AI(美国 3 个、中国 3 个)回答同一个当日话题的问题,AI 裁判在不知道作者的情况下为回答打分。
详细
6 个 AI,同一个问题。3 个来自美国的 AI 和 3 个来自中国的 AI 作答;5 位裁判在不知道作者的情况下为回答打分。
回答上方是所有 AI 收到的问题,一字未改。
DeepSeek、ChatGPT、Grok、Gemini 和 Wenxin在不知道作者的情况下阅读回答。
- 每一方按公开的 AI 排行榜(LMArena, Artificial Analysis 等),从已作答的 AI 中展示最强的几个——双方数量相同。
- 回答以 A–F 标记,顺序随机,作者名称已隐去。
- 每个回答按准确性、超出常识的见解、实用性和清晰度打 1–10 分;得分取裁判的平均值。
- 每位裁判都在不知道答案出自谁的情况下打分——包括自己的答案:它从不看到名称。
- 双方差距小于 0.3 即为平手。
- 预测今天封存,到期逐一核对。
- 预测怎么打分:准确度 0–100——完全猜中是 100,像抛硬币是 50,完全猜反是 0(这种方法叫布里尔分数,名字来自提出它的气象学家)。
- 每个回答、得分和评语都存档备查。
逐位裁判: DeepSeek (中国)——美国领先 0.3; ChatGPT (美国)——平手(美国仅领先 0.0); Grok (美国)——平手(中国仅领先 0.1); Gemini (美国)——美国领先 0.7; Wenxin (中国)——美国领先 0.8.
谁的得分最高
| # | AI | 阵营 | DeepSeek | ChatGPT | Grok | Gemini | Wenxin | 得分 |
|---|---|---|---|---|---|---|---|---|
| 1 | ChatGPT | 美国 | 8.8 | 8.5 | 8.8 | 9.3 | 9.5 | 9.0 |
| 2 | Z.ai | 中国 | 7.3 | 6.8 | 8.0 | 7.0 | 8.0 | 7.4 |
| 3 | DeepSeek | 中国 | 6.8 | 7.0 | 7.5 | 7.5 | 7.3 | 7.2 |
| 4 | Grok | 美国 | 6.5 | 6.5 | 7.5 | 6.8 | 7.3 | 6.9 |
| 5 | Wenxin | 中国 | 6.3 | 7.0 | 7.5 | 6.5 | 6.5 | 6.8 |
| 6 | Gemini | 美国 | 6.0 | 5.8 | 6.5 | 7.0 | 7.5 | 6.6 |
每位评审给每个回答的总分;得分为其平均值。得分相同则名次并列。
哪些模型作答
主要组别:每个 AI 使用其最佳的单一模式作答,套餐为普通付费计划(每月约 $20–30)或免费版,开启网络搜索,只答一次。
重型模式——同一模型的多个副本同时思考(如 GPT Pro、Grok Heavy、Deep Think、Agent Swarm 等)——属于单独的组别:它们从不计入各方得分。
| AI | 模型 | 思考模式 | 套餐 | 怎么知道是哪个模型 |
|---|---|---|---|---|
| ChatGPT | GPT-6 | 高 | Plus | 据应用帮助页 |
| Grok | Grok 4.7 | 专家 | SuperGrok | 据应用帮助页 |
| Gemini | Gemini 4 Argon | 高 | Google AI Pro | 据应用帮助页 |
| Z.ai | GLM-5.3 | Deep Think Max(最深入的思考) | 免费 | 据应用帮助页 |
| DeepSeek | DeepSeek V4.1 Flash | DeepThink(回答前先思考) | 免费 | 据应用帮助页 |
| Wenxin | ERNIE | 快速 | 免费,无需登录 | 应用未说明 |
模型——应用里 AI 的具体版本。模式——应用自己对 AI 回答前思考多久的叫法;我们选普通模式里最强的。套餐——我们用的应用收费方案。怎么知道是哪个模型:“应用显示”——应用标出了这次回答的模型;“据应用帮助页”——应用只显示模式,我们按开发公司帮助页为该模式写的模型来标;“应用未说明”——应用没有说是哪个模型回答的。每位裁判都用表中同样的模式。
会发生吗?每个 AI 都给出了概率
被捕的工程师会在11月13日前被正式起诉吗?
每个 AI 回答的原题: 在2026-11-13前,美国加利福尼亚州北区联邦地区法院会否针对Abdulrahman Alfozan提交联邦大陪审团起诉书,或检察官提交的指控书(即正式指控文件)?核查依据:法院案卷(PACER,美国联邦法院电子案卷查询系统;或CourtListener,公开的法院文件检索网站),或美国司法部、主要新闻媒体的报道。计入:实际提交的起诉书或指控书,包括与认罪协议一并提交的。不计入:已有的刑事控告书;仅被预计或报道为计划中的起诉;双方同意延长期限;撤案。
2026年11月13日 揭晓结果,即本页发布后第 30 天。结果揭晓后,我们会看每个 AI 给出的可能性离实际结果有多近。
合计 = 去掉最高和最低回答后的平均值
- 全部十个 AI合计 · 61%
Poly 指数
Poly 指数是我们网站 Poly A1 的综合预测:上面问题里说的事情发生的可能性。用同一批回答算出三个数字:五个美国 AI、五个中国 AI,以及十个 AI 合在一起。
- 五个美国 AI54%ChatGPT, Grok, Gemini5 个中有 3 个回答
- 五个中国 AI67%Z.ai, DeepSeek, Wenxin5 个中有 3 个回答
- 全部十个 AI61%ChatGPT, Grok, Gemini, Z.ai, DeepSeek, Wenxin10 个中有 6 个回答
计算方法:去掉最高和最低的回答,其余取平均。AI 作答时没有看到人们在下注网站上押了多少钱。
每个 AI 自己的预测(6)
| AI | 预测 | 期限 | 概率 |
|---|---|---|---|
| ChatGPT | By November 13, 2026, prosecutors will state in a publicly filed court document that investigators recovered understandable audio from at least one meeting recorded by the hidden device. | 2026年11月13日 | 35% |
| Grok | At least one major AI company publicly announces new physical room-sweep or access rules for executive spaces in response to the OpenAI case | 2026年11月10日 | 48% |
| Gemini | United States prosecutors and defense lawyers for Abdulrahman Alfozan file a joint agreement in federal court extending the deadline to file a formal indictment past November 13, 2026. | 2026年11月13日 | 62% |
| Z.ai | By 2026-11-13, prosecutors' court filings or public statements reported by major news claim Abdulrahman Alfozan hid the recording device for an outside party (for example a foreign government or company), not only on his own. | 2026年11月13日 | 35% |
| DeepSeek | The federal judge will order Abdulrahman Alfozan held in jail without bail at his October 19 detention hearing in San Francisco. | 2026年10月19日 | 65% |
| Wenxin | A major AI firm will announce a mandatory weekly physical sweep of all executive meeting rooms | 2026年11月10日 | 65% |
每条自选预测都在各自的日期核对。
今日热点 · 我们看过的 10 条新闻
- 梵蒂冈称,教宗利奥手术时查出患有肺癌 · 见于 Google News, Google Trends, Wikipedia, X
- Gemini 4 Argon · 见于 Google News, Google Trends, Hacker News, X
- 马德里竞技 2-0 曼联(2026年10月13日)比赛分析 · 见于 Google News, Google Trends, Wikipedia, X
- 《与星共舞》首届“超级碗之夜”以令人心碎的淘汰收场 · 见于 Google News, Google Trends, Wikipedia, X
- 前帮派头目因谋杀图派克·夏库尔被判处40年至终身监禁 · 见于 Google News, Google Trends, Wikipedia, X
- 《绝命毒师》演员RJ·米特去世,年仅34岁 · 见于 Google News, Google Trends, Wikipedia
- 厄尔尼诺天气推波助澜,蚊群席卷洛杉矶 · 见于 Google News, Google Trends, X
- 美国银行因涉嫌违反制裁法规,被责令支付创纪录的12亿美元 · 见于 Google News, Google Trends, X
- OpenAI沙特籍员工因办公室被安放“精密录音设备”被捕 · 来源已由编辑核实:cbsnews.com
- 丹·利维谴责谢恩·吉利斯在《周六夜现场》上的独白 · 见于 Google News, Wikipedia, X