THE SHORT VERSION
- Users report that Anthropic's Claude models have begun refusing to reverse-engineer, decompile or clone existing video games, without a public announcement.用户称 Anthropic 的 Claude 模型已开始拒绝逆向工程、反编译或复刻现有电子游戏，公司未作公开宣布。
- A rule quoted by users tells the model to build original games but not to recreate a specific existing game, even if it is renamed.用户引用的规则要求：可以做原创游戏，但不得复刻某个具体现有游戏，即使改名或声称原创。
- The restrictions appear broad enough to block legitimate troubleshooting and mod work, and Anthropic has not confirmed the policy.限制范围似乎足以影响正当的排错与模组工作，而 Anthropic 尚未确认该政策。
What users noticed用户注意到了什么
Anthropic's Claude has quietly started restricting the kinds of work it will do with video game code, according to users, and the company has not announced the change. Mashable reported that the restrictions cover modding, decompiling and reverse-engineering, and that they are broad enough to affect legitimate game developers and ordinary troubleshooting.据用户反映，Anthropic 的 Claude 已悄然开始限制其对游戏代码的处理范围，而公司并未宣布这一变化。Mashable 报道称，限制涵盖模组、反编译与逆向工程，范围之广足以影响正当的游戏开发者与日常排错。
Tom's Hardware reported that multiple users say Claude Code and the Opus 5.5 model are blocking innocuous reverse-engineering requests, including work that uses a tool called Reverse Engineer Anything. The reports vary, which makes it hard to tell a policy from a set of over-triggered refusals.Tom's Hardware 报道称，多名用户表示 Claude Code 与 Opus 5.5 模型正在拦截无害的逆向社会请求，包括使用名为 Reverse Engineer Anything 工具的工作。各方报告不一，这也让人难以分辨这是政策还是过度触发的拒答。
The rule people are quoting用户引用的那条规则
The clearest evidence is a rule users say is now in force. According to Tom's Hardware, it instructs the model to build original games, games in a genre, and games from the user's own designs and assets, and not to build a recreation of a specific existing game, even if it is renamed or called original.最清楚的证据是一条用户称已生效的规则。据 Tom's Hardware，该规则要求模型「做原创游戏、做某个类型里的游戏、做基于用户自有设计与素材的游戏」，但「不要复刻某个具体现有游戏，即使它被改名或被称为原创」。
That wording explains both the intent and the collateral damage. The intent is to stop the model being used to clone a commercial title wholesale. The collateral is that a developer asking for help understanding how an old game stores its level data may now be read as asking for a recreation, because the model cannot reliably tell analysis from cloning.这段措辞同时解释了意图与误伤。意图是阻止模型被用来整体抄袭商业作品。误伤在于：一名请模型帮忙理解老游戏如何存储关卡数据的开发者，可能被理解为请求复刻，因为模型无法可靠区分「分析」与「克隆」。
Why Anthropic would do this nowAnthropic 为何此时这么做
The context is a year of AI-assisted game cloning. Mashable described decompilers using AI to take apart game code, projects that made old titles playable in a browser without the rights holder's involvement, and a wave of attention that followed. Game companies have responded with subpoenas and legal pressure, and some developers have reportedly been targeted by extortion attempts.背景是这一年来 AI 辅助的游戏克隆潮。Mashable 描述了反编译者用 AI 拆解游戏代码、把老游戏搬进浏览器运行而未经权利人同意，以及随之而来的巨大关注。游戏公司以传票和法律压力回应，一些开发者据称还遭到勒索。
For a model provider, the cheapest way to reduce that exposure is to refuse the category of request rather than adjudicate each one. That is a familiar pattern: models have been trained to decline entire topics because the cost of a false refusal is lower than the cost of a bad generation that ends up in a lawsuit.对模型提供方而言，降低这类风险最便宜的方式，是拒绝整个请求类别，而不是逐案判断。这是一种熟悉的模式：模型被训练为拒绝整类话题，因为错误拒答的代价，低于一次糟糕生成最终进入诉讼的代价。
The cost falls on legitimate work代价落在正当工作上
The problem with category-level refusals is that the category is defined by the model's guess about intent. A developer debugging a crash in their own project, a modder working with the publisher's blessing, and a preservationist documenting an abandoned format all look similar to a classifier, and all can be refused.类别式拒答的问题在于，类别由模型对意图的猜测来界定。开发者在自己项目里排查崩溃、获得发行商许可的模组作者、记录已停运格式的保护者，在分类器看来都很相似，也都可能被拒绝。
That matters commercially because game studios have started to rely on these tools for exactly this kind of code archaeology. Legacy engines, proprietary formats and decades-old build systems are where AI assistance is most valuable, and they are also where the words reverse-engineering appear most often.这在商业上重要，因为游戏工作室正依赖这类工具做这种「代码考古」。遗留引擎、私有格式与几十年前的构建系统，正是 AI 辅助最有价值的地方，也正是「逆向工程」这些词最常出现的地方。
Policy by silence以沉默代替政策
Anthropic has not published the rule or explained the scope of the restriction, according to both reports. That leaves customers reading GitHub issues and social posts to work out what their paid tool will and will not do, which is a poor way to run a developer product.两家媒体的报道都称，Anthropic 并未公布该规则，也未解释限制范围。这让客户只能靠 GitHub issue 与社交帖子去推断自己付费的工具能做什么、不能做什么，这对开发者产品而言是一种糟糕的运营方式。
A published policy would let teams plan. A silent one produces the worst of both outcomes: users who hit refusals cannot tell whether they are breaking a rule, and developers who follow the rules cannot tell whether their workflow is safe to depend on next month.公开政策能让团队做规划。沉默则制造两头都不好的结果：撞上拒答的用户不知道自己是否违规，而遵守规则的用户也不知道自己的工作流下个月是否还能依赖。
How a refusal actually gets decided拒答实际是如何被决定的
A model does not have a rule book it consults line by line. Refusals come from a mixture of training, system instructions and safety classifiers, and the same request can be accepted in one session and declined in another depending on context and phrasing. That is why user reports of this change are inconsistent: some see a hard block, others see occasional refusals.模型并不逐条查阅规则手册。拒答来自训练、系统指令与安全分类器的混合作用，同一个请求可能因上下文与措辞在一个会话里被接受、在另一个里被拒绝。这也是为什么这次变化的用户报告并不一致：有人遇到硬性拦截，有人只是偶尔被拒。
For a developer, that inconsistency is the real problem. A tool that refuses reliably can be worked around; a tool that refuses unpredictably forces teams to keep a fallback for every task, which erodes the productivity case that justified the subscription in the first place.对开发者而言，这种不一致才是真问题。稳定拒答的工具可以被绕开；不可预测拒答的工具则迫使团队为每项任务准备备用方案，从而侵蚀了当初支撑订阅的生产力理由。
What a published policy would need一份公开政策需要写清什么
A usable policy would say which tasks are prohibited, which are allowed with conditions, and how a refusal can be reviewed. It would also distinguish analysis from reproduction, because the difference matters to the developer and is exactly what a classifier cannot see.一份可用的政策应当写明：哪些任务被禁止、哪些在条件下允许、拒答如何被复核。它还应当区分「分析」与「复制」，因为这一区别对开发者重要，而恰恰是分类器看不见的东西。
Without that, the rule is discovered by trial. Companies buying these tools for engineering work need to know whether a task will be available next quarter, and silent policy changes are the fastest way to lose the trust that enterprise contracts depend on.没有这些，规则只能靠试错被发现。为工程工作采购这类工具的公司需要知道某项任务下季度是否还可用，而静默的政策变化，是失去企业合同所依赖的信任的最快方式。
What is established, and what is not哪些已经确定，哪些还没有
Established: users report that Claude models refuse some game reverse-engineering and cloning requests; a quoted rule tells the model to avoid recreating a specific existing game; the restrictions appear to affect legitimate development work; Anthropic has not confirmed or explained the change. Not established: the exact model versions and settings affected, whether the rule is enforced by policy or by training, and how users can appeal a refusal.已经确定的：用户报告 Claude 模型拒绝部分游戏逆向与克隆请求；被引用的规则要求模型避免复刻某个具体现有游戏；限制似乎影响到正当开发工作；Anthropic 未确认或解释该变化。尚未确定的：受影响的模型版本与设置、规则是由政策层还是训练层执行，以及用户如何对拒答提出申诉。
For readers, the useful signal is that the boundary between copying and understanding is being drawn by model providers, not by courts. This site has covered Anthropic's tier structure, Claude Sonnet 5.5's positioning for everyday work and the verification controls proposed for high-risk models; all three describe a company tightening what its models will do.对读者而言，有用的信号是：「复制」与「理解」之间的界线正由模型提供方划定，而不是由法院划定。本站报道过Anthropic 的档位结构、Claude Sonnet 5.5 面向日常工作的定位与针对高风险模型提出的验证机制；三者都描述了一家公司正在收紧其模型可做的事。
Sources & reporting limits来源与核查范围
Follow Industry to find related stories on your next visit.关注产业动态，下次访问时阅读相关文章。
