151 Apps Collectively Reject Doubao's GUI Agent Operations: AI Phone Assistants Face an Ecosystem Siege

AI phone assistants are experiencing a collective backlash from the app ecosystem. After the public comment period for the Screen Automation Execution Protocol (SAEP) ended, Doubao's phone assistant received explicit rejections from 151 apps for its GUI automated proxy operation capability—with high-frequency apps like WeChat, Taobao, Meituan, Kuaishou, and DeepSeek^Ad all on the list. This is not a simple technical interface dispute, but the first large-scale head-on conflict between AI Agents attempting to take over users' screens and the existing app ecosystem.

GUI Automation: The Core Imagination and Biggest Controversy of AI Phones

The essence of GUI automation is letting AI assistants "see" the screen, "tap" buttons, and "swipe" pages just like humans do. Users only need to say "help me buy a carton of milk on Taobao," and the AI assistant can automatically open the app, search for products, compare prices, and place an order. This is regarded as the core capability that distinguishes AI phones from traditional smartphones, and it is also the flagship selling point of the Doubao phone assistant on devices such as the Nubia M153 and Nubia NaviX Ultra.

But it is precisely this capability that has touched the most sensitive nerve of app developers. According to Jinan Daily, Doubao's disclosed rejection list includes a total of 151 apps, covering categories such as social networking, e-commerce, local life, short video, and AI tools. Mass-market high-frequency apps including WeChat, DeepSeek, Taobao, Kuaishou, and Meituan are all on the list.

It should be clarified that "refusing access" does not mean these apps cannot be installed and used on Doubao phones. Users can still manually open WeChat to chat or browse Taobao to shop. What developers are rejecting is only the Doubao phone assistant's automated proxy operation capability that recognizes and simulates taps and swipes through the GUI. In other words, what the apps are blocking is not users, but the AI assistant as a "middleman."

GUI automation operations refer to a human-computer interaction method in which AI interacts through the graphical user interface (GUI), as distinguished from API (application programming interface) calls. In traditional software integration, apps exchange data and functions through APIs, with clear permission boundaries that are auditable and controllable. GUI automation, by contrast, lets AI recognize interface elements (such as buttons, input boxes, and product cards) through screenshots, then simulate human actions such as clicking, swiping, and typing. The advantage of this approach is that it can operate any app without requiring cooperation from the app vendor, but the cost is that it bypasses all permission control mechanisms set at the application layer, effectively allowing AI to perform operations as a "real human user," making it almost indistinguishable from real human behavior from the app's perspective. This is also the fundamental reason why app vendors view it as a security gray area, while AI phone makers view it as a core competitive advantage.

The SAEP protocol (Screen Automation Execution Protocol) is an industry declaration mechanism established by Doubao to regulate GUI automation operations. Its core design logic is: first publish the protocol content, give app developers sufficient time for notification and choice during the public comment period, and after the comment period ends, developers still retain the right to declare refusal. This "post-hoc opt-out" mechanism is essentially a self-regulatory attempt by an AI vendor in the absence of unified industry standards—hoping to promote the implementation of GUI automation capabilities while trying to ease the app ecosystem's resistance through transparency. But judging from the result of 151 apps collectively exercising their right to refuse, this self-regulatory mechanism did not earn the trust of app vendors; instead, it became a window for the latter to collectively express opposition.

What Are App Vendors Worried About?

From the perspective of app vendors, the risks brought by GUI automated proxy operations are multi-layered.

At the security and risk control level, automated clicks and swipes simulate real user behavior, but their behavioral traces differ from those of humans and may bypass app security verification mechanisms, giving black and gray industries an opening. Apps involving funds and privacy, such as WeChat and Alipay, are especially sensitive to this.

At the business model level, when an AI assistant opens Taobao on behalf of the user and completes a purchase, Taobao's carefully designed recommendation algorithms, ad placements, and promotional pop-ups are all "skipped." Users no longer browse the homepage feed and are no longer influenced by "guess you like," significantly weakening the platform's traffic distribution rights and advertising monetization capability.

At the data sovereignty level, GUI automation means the AI assistant can read all content on the screen—chat records, order information, browsing history. Once such data is collected by the AI assistant and used for model training or cross-app transfer, the user data assets that app vendors have accumulated over the long term will face the risk of being "intercepted."

The design of the SAEP protocol itself gave developers the right to "opt out after the fact"—developers can still declare refusal after the public comment period ends. The collective exercise of this right by 151 apps shows that the industry's resistance to GUI automation is widespread and firm.

Tencent's Green Bubble Short Drama Launches: A Defensive Counterattack Under Traffic Anxiety

In the same time window when AI assistants encountered ecosystem resistance, Tencent quietly launched the Green Bubble Short Drama mini program, introduced by the WeChat short drama content center, with the tagline "Free good dramas, watch with one tap," and a product identity using WeChat brand green.

The direct benchmark for this product is ByteDance's Hongguo Short Drama. According to Jiemian News and National Business Daily, as of July 2026, the standalone Hongguo Short Drama app had 168 million daily active users, up 107% year over year, already exceeding the combined daily active users of the four major long-video platforms Tencent Video, iQIYI, Youku, and Mango TV. Everyone can see the pressure this number exerts on Tencent.

The most prominent feature of Green Bubble Short Drama is not content, but its strong social attribute—the homepage specifically sets up a "Friends Are Watching" recommendation section, completely different from the algorithmic recommendation logic of traditional short-video platforms. This is WeChat ecosystem's differentiated approach: using social relationship chains to provide a new entry point for short drama distribution, rather than fighting head-on on the algorithmic recommendation track where ByteDance excels. Users do not need to download a standalone app and can watch within the mini program, reducing customer acquisition costs.

Tencent partners have recently been applying for the Green Bubble Short Drama trademark, making the productization intent clear. Netizens joked that "watching Hongguoguo and Green Bubble as a child, and still Hongguoguo and Green Bubble as an adult," which aptly summarizes the dramatic nature of this short drama battle.

Mini program short dramas refer to vertical-screen short drama content distributed and played through the mini program capabilities of super apps such as WeChat. Unlike traditional standalone video apps, mini program short dramas do not require users to download and install them; clicking a link allows viewing, greatly lowering the customer acquisition threshold and user decision cost. For WeChat, the strategic value of Green Bubble Short Drama is not only competition at the content level, but also the defense of traffic distribution rights—short drama is currently the content category with the fastest growth in user time spent. If users' habit of watching short dramas is completely occupied by ByteDance products, WeChat's user dwell time and advertising inventory as a super app will be eroded. Choosing social relationship chains ("Friends Are Watching") as the recommendation logic is a differentiated barrier WeChat has built using its unique social data assets, given that its algorithmic recommendation capability does not hold an absolute advantage. Whether this "social + short drama" combination can work will test whether the underlying efficiency of the WeChat ecosystem in content distribution is enough to compete with ByteDance's algorithmic machine.

Hongguo Short Drama Dumbing-Down Controversy: Dopamine Overdraft or Original Sin of Content?

Another side of Hongguo Short Drama has also surfaced. The topic #HongguoShortDramaDumbsDown hit the trending list, with many netizens posting their phone screen time, showing average daily usage of Hongguo Short Drama exceeding 10 hours. Zou Ying, deputy director of the Clinical Psychology Department at the Second Affiliated Hospital of Zhejiang Chinese Medical University, offered a professional explanation: the密集 reversals and hooks in short dramas continuously stimulate dopamine, causing the brain to become passively excited and overloaded. After long-term high-frequency consumption of this neurotransmitter, handling daily tasks becomes easier to fatigue—it is not that intelligence actually declines, but that dopamine reserves are overdrawn, temporarily worsening focus and comprehension.

According to statistics from a third-party data agency, among the top 50 apps by monthly usage time in China in August 2026, ByteDance products accounted for 41.8% of total time, with Hongguo Short Drama at about 125 minutes/day and Douyin at about 112-120 minutes/day. This absolute dominance in user time is precisely the fundamental reason Tencent must enter the short drama market.

The controversy itself is also divided. Some netizens believe short dramas reduce focus and patience, with formulaic plots and logic that cannot withstand scrutiny; other users say watching short dramas is a personal leisure choice, and many AI short dramas are now produced to a relatively refined standard, so there is no need to praise one and bash the other. The boundary between the medical explanation of dopamine overdraft and the moral criticism of "original sin of content" is not clear.

OpenAI Wiretapping Incident and AI Security Acquisition: Trust Crisis and Capability Reinforcement in Parallel

OpenAI recently found itself involved in two security-related incidents at the same time.

According to Kechuangban Daily and AI Pris, Abdulrahman Alfozan, a Saudi national working as an engineer at OpenAI, was arrested on suspicion of hiding an encrypted recording device in the San Francisco office. In late September this year, he wrapped a small black recording device in toilet paper and plugged it into the USB port behind a television in one of OpenAI's conference rooms—and this conference room was precisely where executives held highly confidential technical, strategic, and business discussions. After discovering the device during a routine security inspection, OpenAI did not alert the suspect, but quietly installed a camera in the conference room and captured footage on October 5 of Alfozan sneaking into the room. The device had been in place for about two days before being discovered. Alfozan had worked on OpenAI's product team for more than a year and is currently charged with using or attempting to use a device to intercept oral communications, and could face up to five years in prison. OpenAI said there is currently no indication that he had access to the model weights powering the company's AI tools.

At almost the same time, Sina Finance reported that OpenAI is in preliminary talks to acquire cybersecurity startup Trail of Bits. Founded in 2012, Trail of Bits focuses on discovering and fixing vulnerabilities in open-source software, with a team of more than 40 people, mainly security engineers. In June this year, the company collaborated with OpenAI to launch the "Patch the Planet" project, relying on OpenAI Codex series AI models to find and fix code vulnerabilities in popular open-source projects such as Python.

Viewed together, the two incidents show that OpenAI's security layout presents a dual-track logic of "internal defense and external offense": internally, it needs to plug management loopholes that allow insider leaks; externally, it seeks to fill gaps in the security capabilities of AI models themselves through acquisitions. The security challenges facing AI companies are no longer limited to model alignment; physical security, personnel management, and open-source supply chain vulnerabilities are all becoming new risk surfaces.

Model security and physical security at AI companies are two different types of risk surfaces that are equally important. Model security usually refers to technical protections such as model alignment, jailbreak prevention, and output content compliance; physical security involves traditional corporate security areas such as office access control, hardware device management, and insider behavior auditing. The Alfozan incident exposed a common management blind spot during the rapid expansion phase of AI companies: open office spaces, lax control over hardware device entry and exit, and insufficient physical protection of sensitive meeting locations. Trail of Bits, by contrast, represents another path—systematic vulnerability discovery in the open-source supply chain that AI models depend on through a professional security team. As AI models become increasingly complex and open-source dependencies multiply, security review at a single position is no longer sufficient to address full-chain risks. Acquiring a professional security company means upgrading security capabilities from "point defense" to "systematic engineering."

Other Notable Developments

• Tian Keyu denies model poisoning: After being fired by ByteDance, he founded an AI lab focused on world models, has completed nearly $30 million in financing, and has a post-investment valuation of $200 million. On a podcast, he said that because he could not stand interns arbitrarily stopping other people's programs, he "fought violence with violence" and used GPT to write a program to stop that intern's tasks. But former team colleagues released a report of more than 10,000 words pointing out discrepancies between his account and the facts. (Hongxing Capital Bureau)
• iPhoneDuo completes large model filing: The China version (model A3721) added filing numbers for the large models it carries, involving Apple Intelligence, Tongyi Wanxiang, and Tongyi Qianwen. (Jinan Daily)
• Xpeng opens its smart car technology externally: Covering the vehicle platform, modular technology, and a full toolchain, while also opening the full-domain physical AI ecosystem solutions, including humanoid robots, Robotaxi, flying cars, and Turing AI chips. (Yicai)
• Step Terminal releases its first agent phone STEPX Neo: Priced at 3,999 yuan, equipped with the agent-native operating system Step AOS 1.0 and personal agent Step Amoo. (Caiwen, Lanjinger)
• AMD's Su姿丰 says this year is the "year of AI inference": She expects AI inference to account for about 60% of the market in the future, and the Helios mass-production system has begun deliveries to customers including OpenAI, Meta, Microsoft, Oracle, and Anthropic. (ITHome)
• Microsoft Azure reportedly asked to cut costs by 20%: Girish Bablani, president of the Azure software division, asked about 11,500 subordinate executives to come up with cost reduction plans of at least 20%; whether layoffs are included has not yet been determined. (Sina Finance)

One-line Summary

Doubao's GUI proxy operations were rejected by 151 apps, sparking a head-on conflict between AI phone assistants and the existing app ecosystem.

Paragraph Summary

After the SAEP public comment period ended, Doubao's phone assistant received explicit rejections from 151 apps including WeChat, Taobao, Meituan, Kuaishou, and DeepSeek for its GUI automated proxy operation capability. This is not a simple interface dispute, but the first large-scale conflict between AI Agents attempting to take over users' screens and the app ecosystem. App vendors' concerns center on three levels: security risk control, business models being bypassed by ad skipping, and data sovereignty being intercepted. At the same time, Tencent launched the Green Bubble Short Drama mini program to compete with Hongguo Short Drama, which has 168 million daily active users, using a differentiated social relationship chain approach. The "dumbing down" controversy surrounding Hongguo Short Drama reveals the overdraft effect of short dramas on the dopamine system. On OpenAI's side, an engineer was arrested for hiding a wiretapping device in a conference room, while the company is also negotiating to acquire security vendor Trail of Bits, showing a dual-track security layout of internal defense and external offense. Step Terminal released its first agent phone STEPX Neo, pushing the AI phone track into the system-level agent-native stage.