The Doubao phone challenges the Android ecosystem by simulating user operations with system-level AI permissions, prompting WeChat's ban and raising industry safety concerns.
December 2025 is destined to leave an unusually colorful mark on the history of consumer electronics.
While major phone makers were locked in a year-end sales brawl over specs, the tech world's spotlight unexpectedly landed on an engineering validation unit bearing the model number nubia M153.
Looking only at the hardware spec sheet, this is clearly not a flagship built to win a spec war: it carries the ZTE nubia brand, uses a proven, mature chassis design, and its configuration is decidedly pragmatic. Held in the hand, the unassuming feel of an engineering validation unit seems to tell you silently: in the AI era, the value of hardware lies not in the luxury of its shell, but in the weight of its soul.
Yet this understated device triggered a phenomenal frenzy in the second-hand market, with Xianyu (Goofish) transaction prices at one point surging to 12,000 yuan. And even with cash in hand, you might not be able to "beg" one, because in the eyes of sharp-eyed tech enthusiasts and geeks, it has already transcended the very definition of a phone, being regarded as a "ticket" to a new world.
Image
Crazy? Not at all.
Behind the buying frenzy is a bet on a brand-new possibility: this is the first AI Agent in the Android ecosystem to truly land with system-level highest privileges.
Luo Yonghao shouted on Weibo that "technological revolution cannot be stopped," while Zhou Hongyi predicted that "the mobile operating system landscape is about to be turned upside down." WeChat, meanwhile, made the response most consistent with its security logic: detecting an abnormal environment, it forced users offline.
This isn't just the launch of a phone; it reads more like a declaration of war on interaction paradigms. This device, equipped with "Doubao Phone Assistant," crashed like a disruptive catfish into the app-siloed world we've long taken for granted.
Today, let's skip the boring benchmark scores and talk only about the underlying war underway—the question of who truly owns your phone.
See That "Ghost Finger"?
Why has this phone gotten industry insiders so excited?
The core isn't the hardware—it's that Doubao Phone Assistant. Don't be fooled by the word "assistant"; it's a completely different species from Siri or Xiao Ai.
To understand the gravity of this, we have to get into some hardcore low-level logic.
In Android's existing permission hierarchy, Google has designed a strict sandbox mechanism: App A generally cannot directly control App B. In the past, those plug-ins that grabbed red envelopes or skipped ads typically used AccessibilityService. But such assistive services suffer from limitations including high latency and being easily detected and blocked by app vendors through signature-based identification.
Compared to the high latency and easy blocking often associated with traditional assistive services, Doubao has chosen to directly invoke the system-level System Signature permission. Industry consensus holds that this is aimed at pursuing a more stable and efficient user experience.
With this permission, it is no longer constrained by the application-layer UI tree or accessibility services; instead, it can directly inject events into the Linux kernel's Input Subsystem. Even without the screen lighting up, the system kernel still receives the electrical signals of a finger tap.
In plain terms, this is a ghost finger visible only to code. The click commands it issues are, from the perspective of the underlying input_event structure, nearly indistinguishable from the electrical signal commands generated by a real user's touch on the screen.
That's a dimensionality reduction strike. Previous AI could only look at the screen (CV recognition); the current Doubao can not only see, but also act.
In the past, to buy a carton of milk, you'd open Taobao, search keywords, calculate prices among live streams and ads, claim coupons, and finally place the order. Now, based on LAM logic, you only need to say to your phone: "Buy me the cheapest carton of milk."
The AI will then operate in the background like a seasoned worker: automatically waking up the e-commerce app, automatically comparing prices, automatically bypassing marketing traps, and directly placing the order. In its presence, Taobao or Meituan instantly degrade into a mere product database and logistics interface.
It's worth emphasizing that Doubao Phone Assistant has publicly stated that invoking this INJECT_EVENTS permission requires explicit user authorization and is clearly disclosed in the permission list. As far as the industry understands, current AI assistants all need to use this kind of system-level permission or a similar accessibility permission to provide phone-operating services.
WeChat Had No Choice
Facing this kind of system-level takeover, WeChat reacted fastest and most fiercely—by directly banning it.
Many users who obtained the engineering unit found that once Doubao Assistant was used to send messages via WeChat, it immediately triggered security risk controls, causing the WeChat account to be forcibly logged out.
Behind this lies the core challenge of human-machine verification. All risk control models are fundamentally about confirming that the operator is human. When an AI holds the highest permission to simulate clicks and cannot be isolated by application-layer sandboxes, it becomes, in essence, extremely difficult to distinguish from malicious automated scripts.
For apps like WeChat and Alipay, which involve financial security and core privacy, if the backend cannot tell whether it's the owner or an AI in front of the screen, the entire risk control model faces collapse.
Although Doubao has adopted multiple safeguards (such as requiring manual user takeover for payments and identity verification), once its AI instruction set or cloud-based reasoning system is maliciously exploited, the potential risk is no longer confined to a single app, but could lead to large-scale malicious takeover of user operations—something application-layer risk controls cannot defend against.
This is the deadlock of AI Agent deployment—the chain of trust is broken.
In response to external criticism, Doubao Phone Assistant has also publicly refuted rumors and made adjustments:
On financial security, Doubao has clearly stated that during user-authorized operations, there is no system permission that can bypass passwords, facial recognition, or other authentication steps. All sensitive operations, such as checking bank account balances, still require the user to manually complete the real-person verification and password entry required by the bank's app.
On proactive restrictions, on December 5, Doubao Phone Assistant issued a statement saying it would further restrict its ability to operate banking, internet payment, and other financial apps, and would suspend automated operations involving point-farming or incentive-farming, taking a prudent approach to safeguard financial security.
On proactive communication, Doubao stated it would actively engage with relevant vendors, hoping to jointly define clear and safe AI operation conduct guidelines.
WeChat is guarding its walled garden, but what that wall encloses is not just traffic—it's also the asset security of hundreds of millions of users. The conflict between the two sides is not a simple monopoly dispute, but a generational contradiction between AI trust and traditional risk control.
Apps Are Becoming Plug-ins
Despite the ban and the controversy, that big-mouth Luo Yonghao may have gotten it right this time: this wave cannot be stopped.
He said bluntly on Weibo: "A technological revolution is something no one can hold back, whether the ones trying are big apps or small apps. Whether Doubao Phone succeeds or not, I don't know, but ByteDance's attempt to take this step is truly remarkable. AI assistants will surely bloom everywhere, and our lives will become completely inseparable from them. People in the future will remember this historic day."
These words are not blind cheerleading, but rather they call out a brutal truth: we are standing at the threshold of a major migration in interaction paradigms.
From CLI to GUI to Touch UI to today's LUI (Language User Interface), every transformation has meant a leap in efficiency.
Why did TNT back then die? Because it was artificial stupidity. But today's AI, powered by Transformer architectures and large models, genuinely possesses reasoning and planning capabilities. It can understand vague commands like "buy me the cheapest ticket to Beijing" and break them down into a sequence of operational steps.
More importantly, users are lazy. Laziness is the primary driving force of technological progress.
The wave of de-app-ification is already arriving in the form of Agents. In the future phone, the operating system itself will become the biggest AI, and all apps will be reduced to its back-end plug-ins or API providers.
This also explains why ByteDance launched such a maverick experimental product.
ByteDance missed the earliest land-grab of operating systems. Squeezed between Android and iOS, it has only ever been an app developer. But AI Agent gives it a chance to overtake on a curve. Since it can't build the underlying OS, it will build a super-OS that sits atop the OS.
It's worth noting that Doubao Phone Assistant's current release is a technology preview—an exploratory product aimed at the industry and AI technology enthusiasts, with the goal of gathering feedback and helping form industry consensus and rules.
The nubia M153 may be a rough draft, a special-forces unit assembled from mature hardware, but it has touched the bones of the next-generation OS. In the face of this trend, no matter how big your app is, you either evolve and plug into the AI ecosystem, or you become a relic of a bygone era, waiting to be cleaned up by history.
Handing the Keys to a Black Box?
Don't just cheer. There's a problem that can't be sidestepped.
When we hand over operational control to AI, we also introduce a massive black box.
When your phone allows an AI to simulate clicks, it means the AI holds theoretically unlimited operational authority. And this authority is highly concentrated. Previously, your data was scattered across dozens of apps; now, the system-level AI Agent knows all of it. It has become an omniscient, god-like vantage point.
Even more critically, the current AI large models themselves are black boxes. We don't fully understand the internal reasoning logic of the models.
Regarding user privacy, Doubao has promised that the phone assistant will not store any user screen content in the cloud, that operational processes will not leave storage traces on servers, and that such content will not enter model training. But this commitment relies more on corporate self-discipline than on hard technical constraints.
This concentrated allocation of power still raises an ultimate question about regulation: we cheer that Doubao has broken the monopoly of WeChat and Taobao, returning choice to users. But is it really being returned to users?
No. We have simply taken the operational authority previously scattered across different apps and handed it, in one bundled package, to ByteDance. This brings tremendous convenience, but it also means we've traded it for a new hub of highly concentrated power that is even harder to regulate under traditional frameworks.
In the fully AI-managed era, who regulates this super butler? How do we ensure it doesn't become a super overseer? Liu Xiaochun, director of the Internet Law Research Center at the Chinese Academy of Social Sciences University, stated that for such innovative technologies, an inclusive and prudent attitude is needed—precisely regulating risks while leaving room for innovation. This is precisely the most urgent problem the industry needs to solve.
In Closing
As a technology preview, the nubia M153's mission is already complete.
It tested the boundaries of the Android ecosystem in a nearly radical way, tearing open the walled gardens carefully woven by the giants. It has forced every major company to confront one question: is the future phone a container for apps, or a vehicle for AI?
People in the future will remember this day. Not for this rebranded ZTE phone, but for the fact that it pressed the accelerator on an interaction revolution. It gave us a glimpse of the post-app era—rough, but stunning enough.
As Doubao Phone Assistant itself has urged, safeguarding users' reasonable right to use AI is the red line, but technology serving humanity is the bottom line. Before that day truly arrives, what the industry needs to resolve is how to rebuild a security protocol that people can genuinely trust, even as we break down the silos of apps.
After all, what we want is a capable "Jarvis"—not a "Skynet" that could spiral out of control.