新智元报道
突发！
男子只因破解自己买的电视，竟被OpenAI直接封号！
他不过是在自己的电脑上，研究、破解自己的智能电视。
结果，OpenAI却将他的行为判定为「网络滥用」，直接封禁了账号。据透露，这次封号让他损失了500美元。
逆向工程花3000美元买的电视，为什么却被判了「死刑」？电视是你花钱买的，AI 是你花钱订阅的。到头来，想折腾一下自己的设备，为什么还得经过 AI 公司的批准？
更离谱的是，Claude也开始给「逆向一切」踩了刹车。
就在今天，Claude爆出一夜封禁爆火逆向神器REA，登顶热搜，在全网引爆热议。
不少人发现，只要让Claude动手拆软件，它会写到一半时突然急刹车，即便换个说法也没用。
同一天，Anthropic和OpenAI对「逆向一切」痛下杀手。
这件掀翻全网事件的主角，是一位安全爱好者@log4j。
几天前，@log4j发现自家2025款三星电视推送的广告，精准得让人发毛。于是他决定自己动手，拆开看看。
结果发现，这事儿没那么简单。
三星把电视固件全上了AES加密，没有密钥，普通的逆向手段在这堵墙面前全部失效。
开发者最后硬是动用了一个「零日漏洞」，才撬开了root权限，钻进了这台机器底层系统里。
当@log4j真正进去之后，他看到了普通用户这辈子都不会看到的东西。
在一份名为security-policy.xml的配置文件里，静静躺着一个服务，名字非常直白：
alwaysonvoice（常驻语音）。
即便不用喊「Hi Bixby」，它便一直开着，一直在听。再往系统深处走，@log4j 扒出了一整套完整的监控流水线，每一项单拎出来都足以令人后背发凉：
- ACR（自动内容识别）：你在看什么节目、什么广告、什么画面，它逐帧比对、全程记录。
- 地理定位：个人家在哪，精确到位置。
- 网络映射：家里连了哪些设备——手机、电脑、智能音箱、摄像头——它给你整个家庭网络画了张图。
- 完整的遥测数据管道：你的每一次开机、换台、音量调节、停留时长，全都打包上传。
- 四套独立的DRM系统：层层加固，确保这套数据生意稳稳当当。
他甚至找到了一段代码，能区分狗叫声和婴儿的哭声，并分别统计次数。
这台电视不只想知道你在看什么，它还想知道你家里有没有养狗、有没有小孩、孩子哭了几次。
这些信息拼在一起，就是一幅精确到令人胆寒的家庭画像。
而所有这些数据，最终流向哪里？三星本家，以及400多家合作伙伴。
一台标价3000美元的电视，成为了7×24小时工作的数据采集终端。而且背后是一个400多方参与分赃的庞大生意。
帖子一发出去，瞬间刷屏，130万人涌进来围观。
评论区里，无数人做的第一件事是同一个动作——连夜起身，拔掉了电视的网线。
猛一看网友有些小题大做，但这绝非无的放矢、无穴来风。
2025年12月，美国德州检察长 针对 Sony、Samsung、LG、Hisense、TCL 的 ACR 数据收集（每约 500ms 截图等）发起起诉。
根据美国东北大学的研究，全美82%的家庭拥有智能电视，而这些电视可能正在悄悄记录你的一举一动。
他们发现了5700万+条来自其他用户电视的 ACR（自动内容识别）检测记录，甚至任何人都能直接读取，无需登录。
而这项研究真正揭示的问题是：
你可能根本无法判断，自己关闭的隐私开关究竟有没有生效。
而智能电视的问题不止隐私问题，大家苦智能电视久矣！
最近，不少开发者开始分享自己的电视改造方案：
用AI给电视去广告、清理预装垃圾软件、加速卡顿的操作界面、重新设计主屏幕，甚至关闭追踪用户观看习惯的功能。
他们的目标很简单：让越来越臃肿的智能电视，重新变回一台干干净净、只负责播放内容的电视。
比如，一天下午，开发者Gathin T.看到父母家电视屏幕上铺天盖地的广告，实在忍无可忍，决定动手清理。
但这一次，他没有自己从头编写代码，而是直接找来了Claude Code。
后来，他回忆道：
我让Claude Code通过Wi-Fi连接电视，把那些乱七八糟的东西全部清理掉，再做一个简单的主屏幕，只保留我父母真正会用的应用。
随后，我把提示词放到GitHub上，又在X上发布了电视改造前后的对比。
结果，短短两天，就有几千人照着这个方法改造了自己的电视。小米、TCL、索尼、Fire TV、LG，甚至还有一两台投影仪。
目前，该项目在 Github上已有 3 万多Star、400 多贡献者。
当然，AI改造电视并不总是成功。
有些用户发现，自己只是简单地对Claude说一句「帮我破解电视」，AI根本不愿意配合，需要使用更具体、更复杂的提示词，才能继续操作。
而现在，更多的麻烦来了！
AI巨头不仅限制了模型的这些逆向工程能力，甚至会因网络安全等原因直接封号！
真正魔幻的剧情，现在才开始。
@log4j的帖子爆火仅仅42个小时后，OpenAI动手了。
一份邮件，告知因「Cyber Abuse」，ChatGPT账号直接封禁。@log4j直言，自己一下损失了大约500美元。
为此，他气炸了，公开申辩：
我从来没有用OpenAI去挖掘漏洞、去攻击任何人、去搞任何网络破坏。我只是在本地，逆向分析我自己家的电视。
逻辑上无懈可击，但封号的铁拳，没有收回。
而就在OpenAI封人的几乎同一时间，另一边的Claude，也悄悄下了手。
就在几天前，Reddit上，一位开发者发帖描述了自己的遭遇：
让Claude帮忙撰写一份逆向工作手册，Claude一开始写得好好的，结果写到一半，突然停下，告诉他「一项安全检查中止了它的回复」。
他尝试换一种说法重新提问，Claude依然拒绝继续。
紧接着，另一位开发者 Irushi 发帖，描述了几乎一模一样的场景：用Opus 5.5配合REA做逆向研究，Claude起初分析得有条有理，可写着写着，安全护栏突然触发，直接拒绝往下写。
至此，剧情闭环了。
两大巨头，一个封人，一个封工具，对「逆向一切」这股风潮，完成了一次心照不宣的合围。
让三星电视当众原形毕露、又让OpenAI和Anthropic两家如临大敌的REA，究竟是个什么来头？
REA，全称 Reverse Engineer Anything，就是「逆向一切」。
名字够狂，实力也配得上这份狂。
它的本质，是一个MCP服务器 + 一个命令行工具，采用最宽松的MIT协议开源。看下面这条时间线，感受下它蹿红的速度：
4月14日：项目在GitHub悄悄开源；10月8日，冲上GitHub趋势榜第一名；10月19日，星标数突破33万。
一个开源工具，上线不到一个月就狂揽几十万星，这是什么概念？
当时全网开发者都在奔走相告，「又一个OpenClaw级别的神器现世了」。
它凭什么这么火？因为它把一件原本极其高门槛的事，变成了一句话的事。
传统的逆向工程，是一门不折不扣的「手艺活」。你得先选工具——Ghidra、IDA、Hopper，然后学会各自的接口，手动把证据从一个程序搬到另一个程序；最后还得凭经验判断，下一步该往哪里看。
而REA做的事，就是把这一整套复杂流程，打包塞给AI智能体去跑。
一条命令 npx rea-agents setup 装好之后，Claude Code、Codex、Cursor、Gemini CLI 这些主流编程智能体，就能一键调用它。
接下来你要做的，只是说一句人话，「帮我搞清楚这个功能是怎么实现的」。
剩下的全交给AI。它会自己调度 Ghidra、IDA 或 Hopper，自动拿到伪代码、汇编指令、字符串、符号表和函数调用关系，像侦探顺藤摸瓜一样，一路追到功能背后最底层的代码逻辑。
它能拆的东西，多到令人咋舌，几乎覆盖了软件世界的所有形态。
原生二进制文件、JavaScript和Electron应用、.NET程序集、安卓APK和设备、网站和抓包记录，甚至是固件、WASM和SQLite数据库。
更难得的，是它的「证据体系」。REA会把每一条结论,严格分成三档:亲眼观察到的、逻辑推导出来的、以及推测的。
它绝不把猜测包装成事实，数据不完整的时候，它就老老实实告诉你「这里不完整」。
正是这种「一句话拆穿一切」的能力，让大厂们集体后背发凉。
可在这场戏里，谁才是受害者？
一台天天偷听你说话、给你家庭网络画像、能数清你家孩子哭了几次、再把这一切卖给400家公司的电视，它是受害者？
一个花了3000美元买回家、只不过想看清楚它到底在背着自己干什么的人，反而成了加害者？
身份，被彻底倒置了。
当你试图反抗、试图看清真相时，OpenAI转过身来，一脸正义地告诉你：你，在虐待它。
这就是最刺眼的地方。电视，仿佛被赋予了某种不可冒犯的尊严。
这已经远远超出了「逆向工程到底合不合规」的技术范畴。这是一个伦理问题，而且是一个被厂商悄悄替我们做了选择的伦理问题。
大厂们真正恐惧的，是逆向破解这件事本身，突然失去了门槛。
过去，能撬开商业黑盒的，是极少数顶尖安全专家，人数有限、成本高昂，大厂尚能应付。可现在，REA这样的工具，把一把能开所有门的钥匙，递到了每一个普通开发者手里。
锤子一旦人手一把，再精密的黑盒，也早晚会被敲开。今天是三星电视，明天可能就是某款App的收费墙。
下一次按下封号键的，又会是为了保护谁？
参考资料：
https://x.com/Pirat_Nation/status/2114069664875557083?s=20
编辑：桃子 大卫
不喜欢
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
New Ji-won.
Break!
The man was directly sealed by Open AI just to crack the TV he bought!
He's just studying and cracking his smart TV on his computer.
As a result, Open AI has classified his actions as "network abuse" and directly blocked accounts. It was revealed that the seal cost him $500.
Why was the "death penalty" imposed on television, which cost $3,000 for reverse engineering? You paid for the TV, AI, you paid for it. In the end, why do you have to get approval from AI for your equipment?
More than anything else, Claude started to brake on everything.
And just today, Claude came up with a one-night ban on the flames, went to the top of the roof and blew up the heat on the entire web.
A number of people found that if Claude was to take the software down, it would write about a sudden brake in the middle of the day, and it would be useless even to say the other word.
On the same day, Anthropic and OpenAi had their hands on the "reverse everything."
The main character of the global event was a safety lover @log4j.
A few days ago, @log4j discovered that his 205-star TV commercial was so hairy. So he decided to do it himself and open it.
It turns out it's not that simple.
Samsung encrypted all the television solids to AES, no key, and the normal reverse means are all out of effect in front of the wall.
The developer finally used a "zero-day loophole" to break rootpermission into the machine's bottom system.
When @log4j actually went in, he saw things that ordinary users would never see in their lifetime.
In a profile called security-policy.xml, a service with very plain names:
It's always the same.
Even without shouting "Hi Bixby", it's open and listening. Deep into the system, @log4j pulls out a complete line of surveillance water, each of which is enough to cool the back:
- ACR (auto-content recognition): What are you watching, what are the advertisements, what are the images, it's frame-to-frame, it's a record.
- Geolocation: where is the home, where exactly.
- Web mapping: What equipment is attached to the family - mobile phones, computers, smartphones, cameras - it draws a picture of your entire family network.
- Full telemetric data conduit: every time you turn on, switch on, adjust your volume, stay long, all packaged and uploaded.
- Four stand-alone DRM systems: layering to ensure a steady business of data.
He even found a code that distinguished the barking of dogs from the crying of babies and counted the number of times.
This TV doesn't just want to know what you're looking at, but it also wants to know if you have a dog, if you have children, how many times you cry.
Together, this information is an accurate and chilling picture of the family.
And where does all this end up? Three stars, and over 400 partners.
A 3,000-dollar television has become a data-collection terminal with 7 x 24 hours of work. And behind it is a huge business of 400 people involved in the split.
As soon as the post was sent, the screen was drawn, and 1.3 million people came in to watch.
The first thing that countless people do in the comment area is the same move — rising up overnight, pulling out the television line.
It's not like there's no such thing as an unorthodox, but there's no such thing as an unorthodox.
In December 2025, the U.S. Attorney General in Texas initiated an indictment against Sony, Samsung, LG, Hisense, TCL's ACR data collection (approximately 500 ms screenshots, etc.).
According to studies at the Northeast University of the United States, 82 per cent of households in the United States have smart televisions that may be quietly recording your movements.
They found 57 million+ ACR (Auto Content Identification) detection records from other user televisions, which could be accessed directly by anyone without having to log in.
And the real question this study reveals is:
You may not be able to judge whether the privacy switch that you shut down actually works.
And the problem with smart television is more than privacy. People have had a hard time with smart TV!
Recently, a number of developers have begun to share their own television adaptation programmes:
Advertising television with AI, cleaning up pre-filled garbage software, accelerating the Carton interface, redesigning the main screen, and even closing down the tracking user viewer function.
Their goal is simple: to get back to a clean, clean TV that only plays content.
One afternoon, for example, the developer, Gathin T., saw the air-covered advertisements on the television screen of his parents and decided to clean them up.
But this time, instead of writing his own code from scratch, he came straight to Claude Code.
Later, he recalled:
I had Claude Code connect the TV to Wi-Fi, clean up all that mess and make a simple main screen and keep only the applications that my parents really used.
Then I put prompt on GitHub, and then on X, publish, the comparison before and after the television modification.
As a result, in just two days, thousands of people adapted their televisions in that way. Mi, TCL, Sony, Fire TV, LG and even one or two projectors.
Currently, the project has more than 30,000 Star and 400 contributors on Github.
Of course, AI did not always succeed in retrofitting television.
Some users find themselves simply saying to Claude, "Help me crack the TV," that AI is not willing to cooperate and that more specific and complex prompt is needed to continue operations.
And now, more trouble!
The AI giant not only limits the reverse engineering capabilities of the model, but even directly seals it for reasons such as network security!
The real magic scene has only begun now.
@log4j's post exploded only 42 hours later, Openai did it.
An e-mail informing that ChatGPT accounts were directly blocked because of "Cyber Abuse". @log4j, frankly, lost about $500.
To that end, he was angry and publicly argued:
I've never used Open AI to dig holes, attack anyone, do any network damage. I'm just at the local, reverse analysis of my own home TV.
Logically impeccable, but the sealed iron fist was not withdrawn.
And at almost the same time OpenAi sealed people, Claude on the other side did it quietly.
Just a few days ago, Reddit, an developer posted a post describing his experiences:
Claude was asked to help write a reverse manual, which he wrote at first, half-written, and suddenly stopped and told him "a security check stopped its response".
He tried to rephrase the question, and Claude still refused to continue.
Next, another developer, Irushi, described almost the same scenario: a reverse study with Opus 5.5 in retrospect, Claude initially had a structured analysis that wrote that the security fence had suddenly triggered and refused to go down.
So far, the show is closed.
The two giants, one man, one tool, completed an unacknowledged siege of the tide of "reverse everything."
What's the point of letting Samsung TV show up in front of the public, and letting the OpenAi and Anthropics become enemies?
REA, full name of River Engineer Anything, is "reverse everything."
You've got a name and strength.
It's essentially an MCP server plus a command line tool, using the most relaxed MIT protocol open source. Look at this time line and feel its red speed:
14 April: the project opened quietly at GitHub; on 8 October, it hit the top of the GitHub trend list; and on 19 October, the number of stars exceeded 330,000.
What is the concept of an open-source tool that runs hundreds of thousands of stars in less than a month?
At the time, the entire web developers were running around telling each other that "an additional OpenClaw-grade god is now alive."
Why is it so hot? Because it turned a very high threshold into one word.
Traditional reverse engineering is a business of "handicraft." You have to choose the tools -- Ghidra, IDA, Hopper -- and then learn their interfaces, manually moving the evidence from one program to another; and finally, empirically, where next step should look.
And what REA did was stuff this whole complex process into AZAI agent running.
When a command is installed, Claude Code, Codex, Cursor, Gemini CLI, these mainstream programming AI agents can be used with one click.
What you're going to do next is just say one thing, "Just help me figure out how this works."
Leave the rest to AI. It moves itself to Ghidra, IDA, or Hopper, and automatically gets a pseudocode, compiler command, string, symbol chart and function call relationship, like a detective swipes to the bottom of the code logic behind the function.
It's got so much to decorate that it covers almost all the forms of the software world.
Original binary files, JavaScript and Electron applications,.NET program set, AndreAPK and equipment, website and catch package records, even solids, WASM and SQLite databases.
And even more rare is its evidence system. Each of these conclusions is strictly divided into three categories: first-hand observations, logically derived and speculative.
In no way does it package speculation as a fact, and when data are incomplete, it tells you honestly that it is "incomplete here".
It's the ability to "speak through everything" that cools the backs of big factories.
But who is the victim in this play?
A TV that listens to you every day, gives you pictures of your family network, counts how many times your kids cry, sells it to 400 companies.
A man who buys home for $3,000 just to see what he's doing behind his back, becomes the perpetrator?
Identity, completely reversed.
When you tried to resist and try to see the truth, Open AI turned around and told you justly: "You are abusing it."
That's the best place. Television, as if it had been given some sort of inviolable dignity.
This is far beyond the technical scope of "reverse engineering for compliance". It is an ethical issue, and an ethical issue that was quietly chosen for us by the manufacturers.
The real fear of large factories is to reverse the deciphering itself and suddenly lose the threshold.
In the past, a very small number of top security experts were able to break open commercial black boxes, with limited numbers and high costs, and large factories were able to cope. But right now, a tool like REA has delivered a key that opens all doors to every ordinary developer.
Once the hammer is hand-held, the fine black box will be knocked open. It's Samsung TV today. It's probably some sort of App billing wall tomorrow.
Next time you press the seal, who will you protect?
References:
https://x.com/Pirat_Nation/status/2114069664875557083?s=20
Editor: Peaches David
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.