【文章标题】：Anthropic以”可疑信号”为由封禁了我
【文章正文】： Anthropic banned me for “suspicious signals” Anthropic以”可疑信号”为由封禁了我
I opened my inbox and there it was: Claude, gone. Not a warning. Not a rate limit. A revocation. 当我打开收件箱时，它就在那里：Claude账号消失了。没有警告。没有速率限制。直接撤销。
Anthropic’s Safeguards Team sent me the same template a lot of people seem to be getting lately. “An internal investigation of suspicious signals associated with your account indicates a violation of our Usage Policy.” No clause named. No example. No “here’s what we think you did.” Just: access revoked, appeal by logging back into claude.ai, regards. Anthropic的安全团队给我发送了最近很多人似乎都在收到的模板邮件。“对您账户相关可疑信号的内部调查表明您违反了我们的使用政策。“没有具体条款。没有示例。没有”我们认为您做了什么”。只有：访问权限撤销，请重新登录claude.ai申诉，此致敬礼。
I’ve been a paying Anthropic customer for years. Right now I’m on Claude Max at $200/month. This isn’t a free-tier “maybe they thought I was a bot” story. This is a paid, daily-driver account — the thing I use to actually ship work — taken offline by a sentence that could mean anything. 我作为Anthropic付费用户已有多年。目前使用的是每月200美元的Claude Max套餐。这不是免费层级的”可能他们以为我是机器人”的情况。这是一个付费的日常使用账户——我实际用来交付工作的工具——被一句可能意味任何事的话直接下线。
Wait, so what am I paying for here? 等等，那我到底在为什么付费？
A colleague of mine — completely independent setup, elsewhere in the Philippines — got banned the moment they paid for Max at $100. Same product family. Same sudden wall. I’m on the higher Max tier and still got the template. I keep seeing the same shape next to billing events in other people’s reports. I’m not saying Anthropic admitted the cause. I’m saying “we just took your money” clearly isn’t a shield. 我的一位同事——完全独立的设置，在菲律宾其他地区——在支付100美元升级Max套餐的瞬间就被封禁。同样的产品系列。同样突然的阻碍。我使用的是更高级的Max套餐却依然收到模板邮件。我在其他人的报告里不断看到账单事件后出现的相同模式。我不是说Anthropic承认了原因。我是说”我们刚收了你的钱”显然不是护身符。
I’m not going to pretend Anthropic never has a real enforcement problem. They’ve publicly gone after third-party harnesses riding consumer OAuth — OpenCode and friends showed up in that wave earlier this year. Fine. If you’re abusing the consumer surface as a backdoor API, they will eventually notice. 我不想假装Anthropic从未有过真正的执法问题。他们公开打击过利用消费者OAuth的第三方工具——今年早些时候OpenCode等工具就遭遇了那波清理。这没问题。如果你把消费者界面当作后门API滥用，他们最终会注意到。
See, that is not what “suspicious signals” communicates. Their own Help Center lists the boring categories — repeated Usage Policy hits, unsupported location, ToS. It does not explain what a “suspicious signal” is. The email doesn’t either. The appeal path dumps you into an in-product form after you log back in — no human to plead your case to. You wait. 但问题在于，“可疑信号”这个说法完全没有传达任何信息。他们自己的帮助中心列出了那些乏味的分类——反复违反使用政策、不支持的地区、服务条款。但没有解释什么是”可疑信号”。邮件里也没说明。申诉流程在你重新登录后就把你扔进产品内的表单——没有人工可以申诉。你只能等待。
The nail in the coffin 压垮骆驼的最后一根稻草
I’m not the only one staring at that exact wording. On the Claude Code issue tracker alone: 我不是唯一遭遇这种措辞的人。仅在Claude Code的问题追踪器上：
-
A Brazil marketer hit a mid-checkout “banned organization,” then the same Safeguards email while claiming ordinary coding and content work (#40046).
-
一位巴西营销人员在支付中途遭遇”被禁组织”提示，随后收到同样的安全团队邮件，而他声称只是进行普通编码和内容工作(#40046)
-
A long-running Singapore/Malaysia customer with dozens of clean payments upgraded into Max/Team and watched every related account suspend about eleven hours later — same email (#45936).
-
一位长期使用的新加坡/马来西亚客户，拥有数十次良好支付记录，在升级到Max/Team套餐后约11小时，所有关联账户都被暂停——同样的邮件(#45936)
-
Another $200 Max user on the official CLI described the same March template, then about two months of appeal silence — nobody to talk to (#51670).
-
另一位使用官方CLI的200美元Max用户描述了同样的三月模板邮件，随后经历了约两个月的申诉静默期——无人可沟通(#51670)
Sometimes people get reinstated fast — one founder wrote about a travel-to-India false positive that flipped back in a couple of days. Good for them. That doesn’t make the default experience okay. The default is: template, black box, form, wait. 有时人们会快速恢复——有位创始人写了关于印度旅行误报几天内解决的经历。为他们高兴。但这不能使默认体验变得合理。默认流程是：模板邮件、黑箱操作、表单提交、漫长等待。
Safe AI, trigger-happy bans 安全AI，随意封禁
The irony is not subtle. Anthropic spends a lot of brand equity on being the careful ones — the lab that cares about misuse, the grown-ups in the room. Then the enforcement layer behaves like a 2010s Google account ban: automated, opaque, allergic to context — and disproportionately painful if you are an individual who actually depends on the product. 讽刺毫不隐晦。Anthropic花费大量品牌资产塑造谨慎形象——关心滥用的实验室，房间里靠谱的成年人。然后执行层表现得像2010年代的谷歌账号封禁：自动化、不透明、脱离上下文——如果你是真依赖该产品的个人用户，痛苦会不成比例地放大。
If you are a company burning five figures a month through the API, you have account managers, invoices, and leverage. If you are a person on Max — even $200 Max, even after years of paying them — you have a no-reply address and a reference ID. When that account dies, your productivity dies with it. There is no spare multi-thousand-dollar token budget sitting around to absorb the downtime while Safeguards thinks about your appeal form. 如果你是每月通过API消耗五位数资金的公司，你有客户经理、发票和议价能力。如果你是Max套餐用户——即使是200美元的Max套餐，即使已付费多年——你只有一个无回复地址和一个参考ID。当那个账户死亡时，你的生产力也随之消亡。没有备用的数千美元token预算来消化停机时间，而安全团队还在考虑你的申诉表单。
That’s 2010s Google account-ban energy. Not “we carefully investigated and here is the policy section.” Just: signals. Policy. Revoked. 这就是2010年代谷歌账号封禁的作风。不是”我们仔细调查了，这是相关政策条款”。只有：信号。政策。撤销。
So what’s the point? 那么意义何在？
I filed the appeal. Of course I did. I also stopped trusting them. 我提交了申诉。当然提交了。同时也不再信任他们。
I do not trust that Anthropic will treat a paying individual as a customer when their classifiers get twitchy. I especially do not trust a process where I cannot talk to a human about the work I actually do. And I am done building my week around a single frontier lab that can brick the whole stack with one template email. 我不相信当Anthropic的分类器抽风时，他们会把付费个人当作客户对待。我尤其不信任一个无法让我向真人解释实际工作内容的流程。我已经受够了围绕一个能用模板邮件摧毁整个工作流的尖端实验室来安排我的每周计划。
But man, I already knew this story from the other side of the vault. I previously wrote that Anthropic is Scrooge McDuck — swimming in compute money while nickel-and-diming the people actually using the product. Turns out the vault works both ways: keep all your eggs in it, and one Safeguards morning empties the whole thing. 但老兄，我早从金库的另一边就知道这个故事了。我之前写过Anthropic就像史高治·麦克老鸭——在计算资金里游泳，却对实际使用产品的人斤斤计较。结果证明金库是双向的：把所有鸡蛋放在里面，某个安全团队的早晨就会清空整个金库。
Frontier models are not scarce the way they were two years ago. Other providers are catching up. Open weights out of China keep getting more usable for real work — and diversifying no longer feels like paranoia. It feels like not leaving every egg in Scrooge’s vault. 尖端模型不再像两年前那样稀缺。其他供应商正在追赶。来自中国的开源模型对实际工作越来越可用——多样化不再像是偏执狂行为。这感觉就像是不把所有鸡蛋都放在史高治的金库里。
Anthropic can keep the “safe AI” slogan. I’ll keep a workflow that survives their Safeguards Team having a bad morning. Anthropic可以继续保留”安全AI”的口号。我会保留一个能在他们的安全团队心情不好时存活的工作流程。
Turns out “suspicious signals” is doing a lot of work — just not for the people paying the bill. 事实证明”可疑信号”做了很多工作——只是不是为了付钱的人。