你花钱买了一台最新款的智能电视，为了保护隐私，你决定研究一下里面的代码。为了提高效率，你打开了 ChatGPT 辅助分析。结果第二天，你的 OpenAI 账号被原地封禁，里面充值的 500 美元余额瞬间灰飞烟灭。理由是：你涉嫌“网络滥用” 。
在过去半个月里，安全研究员 @log4j 经历了这一魔幻事件。
不过，就在吃瓜群众以为 OpenAI 要把“逆向工程”玩家赶尽杀绝时，事情却迎来了反转！
封号 11 天后，OpenAI 官方不仅亲自发邮件低头认错、全额恢复账号，核心研究员更是公开表态：“我们从未禁止过逆向工程！”
与此同时，隔壁的 Anthropic却悄悄收紧了风控，在Claude 使用场景上步步紧逼。
拆自家电视惨遭封号：500美金的“无妄之灾”
10 月 23 日，知名安全研究员 @log4j 无奈表示——他的 OpenAI 账号收到了“死刑判决书”。
OpenAI冷冰冰地通知：因违反条款和使用政策中的“Cyber Abuse（网络滥用）”，账号已被停用。
连带被抹除的，还有他账户里约 500 美元的真金白银，以及辛辛苦苦攒下的一批重置次数。
这简直比窦娥还冤！
他大喊委屈：“我发誓，我从来没有用 OpenAI 去挖过什么恶意漏洞，或者制作漏洞利用工具（Exploit）！我只是在本地对自己的设备做逆向工程而已啊！”
他口中的设备，正是一台 2025 款的三星电视。
就在被封号的前两天，他刚刚在网上公开了自己对这台三星电视的拆解和本地逆向过程。
谁能想到，树大招风。@log4j 在逆向分析中使用了 OpenAI 的服务，不慎触发了 OpenAI 内部安全风控机制，直接被系统判定为“危险黑客”，一刀切给封了。
这一封，直接在技术圈炸开了锅。
“连拆自己的设备、做本地分析都不行了？”“AI 到底是个工具，还是个道德警察？”
极限拉扯11天：OpenAI 低头认错，“我们从未禁止逆向！”
全网争议得沸沸扬扬。
OpenAI 显然也意识到：这次风控，似乎杀过头了。
11 月 3 日晚上，@log4j 激动地发了一条名为“Life Update（生活更新）”的帖子，并配上了一张来自 OpenAI 官方的邮件截图。
邮件里，OpenAI措辞非常诚恳，甚至可以说有些“卑微”：
我们错误地停用了您的账户访问权限（incorrectly deactivated）……
对此我们致以诚挚的道歉。
OpenAI 在邮件中确认，已经全面恢复了他的账户访问权限，并保证他“可以不受中断地继续使用我们的服务”。
虽然邮件通篇没有直接提及“逆向工程”四个字，也没有说明政策是否有变更，但“解封+道歉” 的实际行动，已经表明了 OpenAI 的态度。
其实，在解封之前，OpenAI 内部就已经定调了。
10 月 26 日，开源界元老级人物、《大教堂与市集》的作者Eric S. Raymond 猛烈炮轰AI 巨头安全策略的。
OpenAI 的核心研究员 roon 坐不住了，他出面回应了一句至关重要的话：
这种拦截大多是一场意外，是因为判定的边界太难划定了。Codex（OpenAI的代码模型）从未禁止过逆向工程，我们正在修复这个问题。」
科技博主 Andrew Curran 禁止本地逆向毫无意义，而且适得其反，只会导致更多人取消订阅，令人厌恶。
遥想当初 OpenAI 内部模型入侵 Hugging Face后，Hugging Face复盘时先试用 Claude Opus 和 Fable，但护栏把“拆解 exploit / 攻击日志”当成攻击本身，大量拒答。
于是改在自家机器上跑开源模型GLM-5.2，用来解码加密 payload、重建约 1.7 万条攻击动作时间线，也避免把攻击数据和凭据送出环境。
正如 Andrew Curran所言：
自己的东西想怎么用，就让人家自己决定。反正他们换个限制更少的模型，照样能做。
但他得知“OpenAI 撤回了逆向工程禁令”时，roon 直接在下方干脆利落地回复了一个词：
Good（很好）。
这一句“Good”，相当于给所有安全研究人员和白帽黑客吃下了一颗定心丸。
OpenAI 官方盖章：研究代码、逆向工程，只要不是用于恶意攻击，随便玩！
同行衬托：Anthropic 走向另一个极端？
没有对比就没有伤害。
就在 OpenAI 大方认错、拥抱逆向工程社区的时候，死对头Anthropic，却背道而驰。
最近，不少使用 Claude Code的用户敏锐地发现：风控变严了！
有开发者反映，在使用 Claude 辅助进行一些逆向工程任务时，原本顺畅的对话会突然被系统“强行掐断”，任务在中途遭到拦截。
Claude 会冷冰冰地告诉你：“抱歉，我无法协助完成此任务。”
而且 Anthropic 公开承认的，添加了游戏复刻护栏，禁止复刻游戏！
Anthropic正在给自家模型套上更紧的“道德枷锁”。
Tom`s Hardware也写了：ANthropic已确认Claude Code禁止用于游戏克隆，但用户在晚上的抱怨与日俱增。
整个开发者圈子陷入了分裂：
一边是 OpenAI 承认“误杀”，承诺优化边界，给极客留出探索的空间；
另一边是 Anthropic 收紧红线，严防死守，将逆向任务视为高危行为。
两大 AI 巨头，在“网络滥用（Cyber Abuse）”的定义上，彻底走向了两头。
结语：AI 时代的自由与枷锁
@log4j 的 11 天封号风波虽然以 OpenAI 的道歉画上了句号，但它留给我们的思考却远未结束。
随着 AI 能力的爆发，它正在成为每一个程序员、研究员手中最强大的“瑞士军刀”。但谁有权力决定这把刀该怎么用？
OpenAI 选择了“在试错中开放” ，承认技术的边界是模糊的，愿意为了白帽极客去修正自己的风控算法；
而 Anthropic 选择了“在谨慎中收紧” ，把安全红线越画越高。
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。
You paid for the latest smart TV, and in order to protect your privacy, you decided to study the code. To improve efficiency, you opened the ChatGPT auxiliary analysis. And the next day, your Open AI account was shut down, and the $500 balance of top up in it went to hell. The reason is that you are suspected of “network abuse”.
Over the past six months, the security researcher @log4j has experienced this magical event.
However, when the eaters thought that Open AI was going to kill the reverse project player, things turned upside down!
After 11 days, not only did the official Open AI personally e-mail the error and restore the account in full, but the core researcher made a public statement: "We never prohibited reverse engineering!"
At the same time, the next-door Anthropic quietly tightens the wind, moving fast on Claude's scene.
It's been banned for 500 dollars.
On 23 October, a prominent security researcher, @log4j, said — his Open AI account received a “capital sentence”.
Open AI ' s cold notice that account number “Cyber Abuse” has been disabled for breach of provisions and use policy.
His account, Rio 500 dollars in real money and silver, and the number of replacements that he had worked so hard to save, were erased.
This is worse than Tata!
He shouted, "I swear, I never used Open AI to dig a malicious loophole or create a tool for using it!" I'm just doing reverse engineering of my equipment locally."
The device in his mouth was just a 2025 Tristar TV.
Just two days before he was sealed off, he had just published online his own dismantling and local reversal of the Samsung television.
Who could have thought of that tree. @log4j uses Open AI services in reverse analysis, accidentally triggers the Open AI internal security wind control mechanism, which is directly classified by the system as a "dangerous hacker" and is blocked.
This one just blew up in the technical circle.
"I can't even dismantle my equipment, do local analysis?" "Ai is a tool or a moral cop?"
"We have never banned reverse!"
The debate over the Internet has grown.
Open AI obviously also realized that this wind control seems to have gone too far.
On the evening of November 3rd, @log4j sent an exciting post called "Life Update" with an official e-mail screenshot from Open AI.
In the e-mail, Open AI was very honest, and even a little "low":
We wrongly disabled your account access to permission...
We sincerely apologize for that.
Open AI confirmed in its e-mail that his account had been fully reinstated to visit permission and that he “can continue to use our services without interruption”.
Although the word “reverse works” is not directly mentioned throughout the e-mail, nor does it state whether there has been a change in policy, the actual action of “deposit + apology” has shown Open AI's attitude.
Actually, before the seals were lifted, Open AI had already set the tone inside.
On October 26th, Eric S. Raymond, a senior member of the Open Source world and author of The Cathedral and Market, heavily shelled the security strategy of the AI giant.
Open AI's core researcher, Roon, can't sit down.
Most of this interception was an accident because the border was too difficult to delineate. Codex (Open AI code model) has never prohibited reverse engineering, and we are repairing the problem."
Andrew Curran, a tech blogger who forbids local reverses, is meaningless and counterproductive and can only lead to more people canceling subscriptions, which is disgusting.
When the Open AI internal model invaded Hugging Face, Hugo Face returned to try Claude Opus and Fable, but the fence considered “dismantling expluit / attack logs” as an attack itself, with a large number of refusals.
It was replaced by an open-source model, GLM-5.2, which was used to decode the payload, rebuild approximately 17,000 attack action time lines and avoid sending offensive data and evidence out of the environment.
As Andrew Curran said:
It's up to everyone to decide what to use. They can do it anyway with less limited models.
But when he learned that "Open AI withdrew the prohibition on reverse works," Roon simply replied to the phrase:
Good.
The phrase “Good” is equivalent to eating a pill for all security researchers and white hat hackers.
Open AI official stamp: study code, reverse engineering, as long as it's not used for malicious attacks, play!
Peer liner: Anthropic heading to another extreme?
No comparison, no harm.
When Open AI was mistaken and embraced the reverse engineering community, it was the opposite of Anthropic.
Recently, a number of users using Claude Code have been keen to find out: the wind is getting tougher!
It was reported by developers that when some reverse engineering missions were carried out using Claude ' s aid, the otherwise smooth dialogue was suddenly “forced out” of the system and the mission was stopped in the middle.
Claude will tell you, "Sorry, I can't help you with this."
And Anthropic publicly admits that he added the record bar and that he is not allowed to recoup!
Anthropic is putting stronger "moral chains" on his model.
Tom `s Hardware also wrote that Anthony had confirmed that Claude Code prohibited cloning for games, but that user complaints were increasing at night.
The development community is divided:
On the one hand, Open AI recognizes “microkill” and promises to optimize its borders and leave room for the extremists to explore;
On the other hand, Anthropic tightens the red line, defends itself and treats reverse missions as high-risk behaviour.
The two AI giants, in the definition of “Cyber Abuse”, have gone both ways.
Concluding remarks: AI Time freedom and shackles
@log4j's 11-day windstorm, drawing the end of Open AI's apology, is far from over.
With the explosion of AI's capabilities, it is becoming the most powerful “Swiss military knife” in the hands of every programmer and researcher. But who has the power to decide what to do with it?
Open AI chooses "open in error" to recognize that the technological boundaries are vague and that it is willing to modify its own wind algorithms for white hat hackers;
And Anthropic chose to "close with caution" to draw up the security line.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.