11-19 更新 实时
首页/ 科技公司/ 正文

AI 驱动的逆向工程会终结专有软件吗

科技公司 来源:Analytics India Magazine 2026-11-10 19:30 0 阅读 机器翻译
AI 驱动的逆向工程会终结专有软件吗

PayPal 软件工程师、在 Instagram 上拥有超过 280 万粉丝的科技内容创作者 Nishant Tiwari 在着手研究 Uber 的定价机制时,使用了一套 AI 辅助的逆向工程工作流来检查该公司的 Web 应用。他对 Uber 客户端代码的分析识别出了一些脚本,这些脚本似乎会访问电池状态及其他设备信息,包括浏览器和硬件属性。

Nishant 在四部手机上测试了同一段行程——三部 Android 设备 and 一部 iPhone。他将实验重复了五到六次,随后意识到结果显示电池电量与车费之间存在相关性。

这些发现得益于 Reverse Engineering Anything(REA)等工具,这是一个开源 GitHub 项目,将 AI 智能体与逆向工程工具连接起来,帮助开发者检查软件并推断其工作原理。REA 由 GitHub 用户 morluto 创建,后者拒绝参与本篇报道。

该项目在 GitHub 上已获得超过 447,000 个星标——而 10 月初仅为 10,000 个星标——跻身该平台星标数最多的前五大代码库之列。

对于销售专有软件的公司而言,问题在于客户能够从可获取的代码和行为中重建出多少产品功能。

逆向工程是指检查一个已完成的软件产品,以了解其工作原理。开发者可能会检查编译后的二进制文件,使用调试器观察其执行过程,或者分析应用程序如何响应不同的输入。

安全研究员兼工程师 Himanshu Anand 解释说,这一过程传统上需要相当高的技术专长。工程师必须解读汇编指令,并把各个函数连接起来,才能理解更大的系统。大语言模型降低了这一门槛。

“逆向工程一直存在。人们过去用它来调试代码或找出漏洞,但有了 LLM,逆向工程的门槛被降低了,”Anand 对 AIM 表示。

他估计,过去需要一个月完成的逆向工程任务,如今在 AI 辅助下可能只需“一两个小时”。

AI 辅助分析可以帮助开发者复现单个软件功能,尤其是当这些功能的行为可以通过客户端应用观察到时。不过,复现一个功能并不一定意味着能复制整个产品。

桌面或移动应用可能会向用户暴露部分代码,而核心处理则在远程服务器上进行。Anand 以 Gmail 等服务为例指出,用户无法看到完整代码,只能进行黑盒测试,他形容这种方式“非常有限”。

技术咨询公司 The Judge Group 的印度区总裁兼全球交付主管 Abhishek Agarwal 表示,AI 可以帮助开发者识别相关函数,减少在错误线索上耗费的时间。但“产品的许多行为可能依赖于一项私有服务。”

复现出来的功能也可能缺少原产品的更新与支持。企业在购买软件本身的同时,还要为这些服务付费。

然而,也有人担心,AI 驱动的逆向工程工具的普及,可能会为我们所熟知的软件敲响丧钟。据 Tom's Hardware 的一篇报道,Anthropic 已悄然开始限制 Claude 在游戏模组制作、反编译或逆向工程方面的能力。

不过,Anand 认为这些担忧毫无根据,不予认同。“这就像说微软因为开源而完蛋了一样,”他打趣道。

如果另一家厂商更快地做出一个足以令人信服的替代品,而客户又看不出多大差别,那么重金投入某一专有功能的公司就可能面临压力。不过,Agarwal 提醒不要把所有模仿都当作商业威胁。

“熟悉的界面容易被认出来;可靠的集成、客户专属数据和支持则更难复制,”他指出。

Agarwal 建议在相同条件下,把传统逆向工程与 AI 辅助分析进行对比。可以给两个工程团队同样的产品访问权限,要求他们重建某一功能,然后评估结果的可靠性以及所需的人力投入。

“仅仅速度快只是有意思;可重复的商业替代才是更强的信号,”他表示。

CMS INDUSLAW 合伙人 Bharadwaj Jaishankar-Bharadwaj 指出,版权将软件代码作为文学作品予以保护。他还谈到软件前端布局的保护问题,并将其与底层功能区分开来。“版权法的基本原则是,思想不受版权保护,受保护的只是表达,”他说。

然而,律所 Anand & Anand 诉讼业务合伙人 Shantanu Sahay 认为,重现某一功能并不必然构成版权侵权。

如果原作品受保护的实质性要素被再现,那么没有复制源代码这一点并不具有决定性。这一法律问题可能取决于复制了什么,以及这些要素是否受保护。

《版权法》第 52(1)(ac) 条允许为确定计算机程序的内在思想与原理而观察、研究或测试该程序,同时实施该程序所提供功能所必需的行为。第 52(1)(ab) 条涵盖为获取独立创作程序的互操作性所必需的信息而实施的行为,但须满足若干条件,包括合法持有该程序,以及该信息并非以其他方式易于获得。

这些条款并未提供复制或商业性再现软件的概括性许可。Sahay 指出,许可协议和保密协议也可能限制反编译。对闭源软件的保护还取决于其许可协议是如何拟定的。

对企业而言,安全问题始于向客户分发了什么。Agarwal 表示,公司应把客户端应用和二进制文件视为可被检视的对象,而不是仅依赖源代码未公开这一事实。

混淆和防篡改措施可以提高分析难度,但也许只是推迟分析。公司还需要检查其应用包含哪些内容,特别是凭证、敏感规则以及本可保留在服务器上的信息。

Agarwal 建议,在可行的情况下将核心业务逻辑从分发的客户端移出,从软件构建中移除机密信息,并收紧 API 边界。监测有助于发现反复试探,而发布审查应包含对二进制文件、软件包和前端的检查。

他还建议,设计假设应发生转变:不再问某人能否对一款产品进行逆向工程,而是考虑当 AI 智能体能够反复且低成本地检视该产品时会发生什么。

因此,对企业软件公司而言,风险不仅仅在于 AI 能否检视其应用程序,而在于它能否为客户愿意付费的某项功能提供可靠的替代方案。

译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。

查看英文原文(Is AI-Powered Reverse Engineering the Death of Proprietary Software?)

When Nishant Tiwari, a software engineer at PayPal and a tech content creator with over 2.8 million followers on Instagram, set out to examine Uber's pricing, he used an AI-assisted reverse-engineering workflow to inspect the company's web application. His analysis of Uber's client-side code identified scripts that appeared to access battery status and other device information, including browser and hardware attributes.

Nishant tested the same ride across four phones—three Android devices and an iPhone. He repeated the experiment five to six times and realised the results showed a correlation between battery levels and fares.

The findings were enabled by tools such as Reverse Engineering Anything (REA), an open-source GitHub project that connects AI agents to reverse-engineering tools that can help developers examine software and infer how it works. REA was created by GitHub user morluto, who declined to participate in the story.

The project has over 447,000 stars on GitHub—up from just 10,000 stars in early October—listing among the top five most-starred repositories on the platform.

For companies selling proprietary software, the question is how much of their product's functionality customers can reconstruct from the code and behaviour available to them.

Reverse Engineering is Becoming Accessible

Reverse engineering involves examining a finished software product to understand how it works. Developers may inspect a compiled binary, use a debugger to observe its execution, or analyse how an application responds to different inputs.

Security researcher and engineer Himanshu Anand explains the process traditionally required considerable technical expertise. Engineers needed to interpret assembly instructions and connect individual functions to understand the larger system. Large language models have lowered that barrier. 

"Reverse engineering was always there. People used to debug code or find bugs with it, but with LLMs, it has lowered the threshold of reverse engineering," Anand tells AIM.

He estimates that a reverse-engineering task that previously took a month could now take "one or two hours" with AI assistance. 

Limitations

AI-assisted analysis can help a developer recreate individual software features, particularly when their behaviour is visible through a client application. However, reconstructing a feature does not necessarily mean reproducing an entire product.

A desktop or mobile application may expose some of its code to the user, while core processing happens on remote servers. Anand points to services such as Gmail, where users cannot see the full code and can only conduct black-box testing, which he describes as "very limited."

Abhishek Agarwal, President of Judge India and Global Delivery at tech consulting firm The Judge Group, says AI can help developers identify relevant functions and reduce time spent following false leads. But "much of a product's behaviour may depend on a private service."

A recreated feature may also lack the original product's updates and support. Companies pay for these services alongside the software itself.

Yet, there have been concerns that the democratisation of AI-powered reverse engineering tools may sound the death knell for software as we know it. According to a report from Tom’s Hardware, Anthropic has quietly started restricting Claude’s capabilities in modding, decompiling, or reverse-engineering video games. 

However, Anand dismisses those concerns as unfounded. "It’s like saying Microsoft is dead because of open source," he quips.

A company investing heavily in a proprietary feature could face pressure if another vendor builds a convincing substitute sooner and customers see little difference. However, Agarwal cautions against treating every imitation as a commercial threat.

"A familiar screen is easy to recognise; dependable integrations, customer-specific data, and support are harder to reproduce," he notes.

Agarwal suggests comparing conventional reverse engineering with AI-assisted analysis under similar conditions. Two engineering teams could be given the same access to a product and asked to reconstruct a feature, with the results assessed for reliability and the human effort required.

"Speed alone is interesting; repeatable commercial substitution is the stronger signal," he remarks.

What Indian Law Protects

Bharadwaj Jaishankar-Bharadwaj, Partner at CMS INDUSLAW, notes copyright protects software code as a literary work. He also discusses protection for software front end's layout, while distinguishing it from the underlying functionality. "The basic tenet of copyright law is that there is no copyright in an idea; it's only in the expression," he says.

However, Shantanu Sahay, Partner, Litigation, at law firm Anand & Anand, believes recreating a feature does not automatically infringe copyright. 

The absence of copied source code is not conclusive if substantial protectable elements of the original are reproduced. The legal question can depend on what has been copied and whether those elements are protected.

Section 52(1)(ac) of the Copyright Act permits the observation, study, or testing of a computer program to determine its underlying ideas and principles while performing acts necessary for the functions for which it was supplied. Section 52(1)(ab) covers acts necessary to obtain information essential for interoperability of an independently created program, subject to conditions including lawful possession and the information not being otherwise readily available.

These provisions do not provide a blanket permission to copy or commercially reproduce software. Sahay notes licence and non-disclosure agreements can also restrict decompilation. The protection for closed-source software also depends on how their license is crafted.

Limiting Exposure

For enterprises, the security question begins with what is distributed to customers. Agarwal says companies should treat client applications and binaries as inspectable, rather than relying only on the fact that source code is unpublished.

Obfuscation and anti-tampering measures can make analysis harder, but they may only delay it. Companies also need to examine what their applications contain, particularly credentials, sensitive rules, and information that could be kept on the server.

Agarwal recommends moving core business logic away from distributed clients where practical, removing secrets from software builds and tightening API boundaries. Monitoring can help detect repeated probing, while release reviews should include inspection of binaries, packages, and front ends.

He also recommends that the design assumption should change from asking whether someone can reverse engineer a product to considering what happens when an AI agent can inspect it repeatedly and cheaply.

For enterprise software companies, the risk is therefore not simply whether AI can inspect their applications. It is whether it can produce a reliable substitute for a feature customers are willing to pay for.


正文保留英文原文,标题与摘要由机器翻译。

本文由采集器自动同步

原文发布于 Analytics India Magazine(Is AI-Powered Reverse Engineering the Death of Proprietary Software?),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。

阅读原文 ↗