AI is starting to take over mobile phones, and for safety reasons, Baidu has locked 'Lobster' into the cloud
Baker Street Detective
Before letting AI do work on your behalf, it’s essential to first establish secure boundaries
Author：Bao Kemo
On March 12, Baidu released an application in the Android market that seemed rather 'extraordinary' –Red Finger Operator. It is referred to asthe world’s first mobile-based 'lobster' (OpenClaw) application. Simply put:Download an app, and you can 'raise an AI lobster' that automatically performs tasks for you on your phone。
At around 11 AM on the 12th,Founder of OpenClawPeter Steinberger posted two consecutive replies on overseas social media, praising the speed of China's AI innovation as 'Amazing'and expressed willingness to collaborate with Baidu to develop the lobster project, and Baidu quickly responded, expressing anticipation for further collaboration.
What made Baidu respond so quickly and show great interest in this small AI tool developed by an individual? The answer lies in the Operator software launched by Baidu called Red FingerRed Finger Operator software. Prior to this, the industry generally believed that only AI models deeply integrated into underlying systems could achieve functions like 'hailing rides, ordering food delivery, searching for flight tickets, automatic order placement, cross-app operations, and task automation.' Now, the emergence of Red Finger breaks this mindset, allowing automated task execution to have broader use cases.
At this point, some netizens may wonder: Could a program capable of executing tasks automatically be unsafe? Coincidentally, Baidu also considered this issue and provided a solution leveraging its own strengths.Moving the task of 'AI operating mobile phones' to the cloud for completion is highly likely to reduce the risk of misoperations potentially caused by OpenClaw.。
The final piece of the puzzle for large-scale implementation of the OpenClaw application.
Before the emergence of Red Finger Operator, most people believed that the most direct way for an AI Agent to help users operate their phones was to take over device permissions. This implies that, theoretically, with enough permissions, it could operate any app on the phone, thus posing a security risk. Whoever addresses these potential safety issues first will hold the final piece of the puzzle for widespread OpenClaw adoption.
In fact,OpenClaw does indeed carry some risks.Experts from the Ministry of Industry and Information Technology have also publicly stated that when using OpenClaw, six points must be noted: use the latest version; strictly control internet exposure; adhere to the principle of least privilege; cautiously use the skills marketplace; guard against social engineering attacks and browser hijacking; and establish long-term protective mechanisms.
However, under overwhelming promotional campaigns, these six reminders still couldn't suppress the curiosity of ordinary people who wanted to 'raise lobsters.' Multiple large internet companies, through the 'Lobster Market,' helped ordinary users set up the necessary operational environment for raising lobsters. Many reported after installation that they couldn’t afford the essential tokens and struggled to configure the permissions required for lobsters.
In fact, this situation is not difficult to understand. The threshold for installing and using OpenClaw is higher than most people imagine. Even Meta’s AI security researcher, Summer Yue, when using OpenClaw to organize emails, explicitly instructed it to 'only analyze emails and provide suggestions, without making any actions without approval.'
However, during subsequent operations, OpenClaw suddenly began deleting emails on a large scale. Multiple 'stop' commands sent by the user were ineffective, and ultimately, the user had to manually terminate the process to stop the deletion, resulting in hundreds of emails being deleted.
Against this backdrop, to launch an OpenClaw APP for ordinary users, it is crucial to ensure the APP uses minimal permissions while providing essential functions. After much deliberation, Baidu’s solution is to move the task of 'AI operating mobile phones' to the cloud for execution.
Baidu's 'safety valve' for raising lobsters.
Behind Red Finger Operator is actually acloud-based virtual phone, which also serves as the 'safety valve' for regular users to farm lobsters.
OpenClaw does not run on the user's actual phone but is pre-installed in a cloud-based virtual device. After the user sends an instruction, the AI executes the operation within the cloud-based phone, completes the task process, and then returns the result to the user. This ensures that the AI can function normally, but its working platform is not the user’s device, although it can still feed results back to the user’s device.
This architecture brings about a crucial change — the execution environment of the AI is systemically isolated from the user’s real device. Even if the AI needs to perform cross-app operations or execute complex automated tasks, it only runs on the cloud-based virtual system, rather than on the user’s local device.
This means the AI’s permissions are controlled within a manageable and secure boundary.
In terms of technical architecture, Baidu has split OpenClaw and Operator into two different capability layers.
OpenClaw handles complex tasks, such as executing automation processes in PC and web environments: gathering hot topics across the internet to generate reports, searching for resources across web pages and downloading them automatically, performing deep data scraping, etc. Meanwhile, Operator focuses on mobile devices, completing cross-app operations in native app environments, such as hailing rides, ordering food delivery, social interactions, and even idling in games.
Once the two work together,the AI doesn’t just 'answer questions' but begins to truly execute tasks on behalf of the user.。
For instance, users only need to input a single sentence: 'Help me search for the cheapest airfare from Beijing to Changsha on February 18,' and the AI can automatically complete the search, filter results, enter the booking page, and prepare to place an order, leaving only the final confirmation to the user. AI agents are evolving from 'being able to chat' to 'being able to perform tasks,' which is a change happening across the entire industry. And Baidu has clearly already entered this space.
Baidu is one of the earliest tech companies in the country to systematically develop the OpenClaw ecosystem. Just before the Lunar New Year, Baidu’s intelligent cloud introduced a simplified version of the OpenClaw deployment process, encapsulating capabilities like Baidu AI Search, Encyclopedia, e-commerce, and Qianfan Deep Research Agent into Skills, and uploading them to the open community ClawHub, making it easier for developers to invoke them.
At the same time, to address the cost concerns that developers care about most, Baidu also launched the Qianfan Coding Plan subscription, reducing Token usage costs through a monthly subscription model.
At the product entry level,the Baidu app already supports one-click invocation of OpenClaw.After completing deployment, users simply need to click in the search bar or message center to activate this AI assistant.
On March 11,Baidu Intelligent Cloud launched DuClaw Zero Deployment Service,which further simplifies the original one-click deployment to 'zero deployment.' Users don’t need to configure servers, set up large model API Keys, or prepare a runtime environment, allowing them to directly start using the service right away.
From cloud architecture, Skill ecosystems, to mobile entry points, what Baidu is doing is transforming OpenClaw from a developer tool into an AI execution system that ordinary users can also utilize.
If the large models of the past addressed the question of 'whether AI can think,' the current issue has now evolved into 'can AI do work for humans.'
And when AI truly begins to work on behalf of humans, new questions arise: where will it operate? How much can it handle? Who will ensure its safety? Baidu’s answer is to let AI work, but keep it caged in the cloud.
Because when AI starts operating your phone, the stronger its capabilities, the more crucial its boundaries become. And this, precisely, is the true watershed for OpenClaw, rather than how many ordinary people’s devices are equipped with OpenClaw.
Section 03: The Confidence Behind Baidu Launching Red Finger Operator
Many people’s first reaction upon hearing 'running OpenClaw in the cloud' is often 'isn’t that just moving to the cloud? Nowadays, almost all internet services are cloud-based, which doesn’t sound particularly complex.' In fact, enabling an AI Agent to truly run in the cloud isn’t as simple as transferring a program to a server.
OpenClaw is essentially an AI execution system capable of actively operating within a software environment. It not only needs to invoke large models but also continuously execute task flows, interact across applications, handle web and app interactions, and even dynamically adjust execution strategies midway through a task based on environmental changes. This means that the cloud must provide not just computing power, but also a fully operational 'digital device environment.'
In other words,What Baidu aims to achieve is not merely the deployment of an AI model; rather, it seeks to address several complex issues simultaneously in the cloud.。
First isVirtual Device Environment—constructing a complete virtual mobile system within servers, allowing AI to click, navigate, and perform tasks just like it would on a real phone.
Next isComputing Power and Task Scheduling, AI Agent tasks are often long processes with multiple steps, requiring stable and continuous computational power support, while also ensuring resource allocation when a large number of users are using the system simultaneously.
The third isSecurity isolation, AI has cross-application operation capabilities, and without strict isolation mechanisms, it can easily infringe on user data and account security boundaries.
Finally,Low-barrier user experience—ordinary users cannot understand image deployment, API configuration, and computational power management; the system must hide all these complex processes in the background.
These four aspects combined essentially make this not just an AI product issue but a cloud computing systems engineering challenge. Precisely because of this, there are actually few companies that can run OpenClaw-like AI Agent products stably on the cloud.
Baidu’s ability to complete this cloud computing systems engineering project is mainly due to years of accumulated expertise. According to Baidu's Q4 2025 earnings report, its AI cloud infrastructure revenue grew by 34% year-over-year in 2025, and subscription revenue for AI accelerator infrastructure in Q4 2025 surged 143% year-over-year. Corporate demand for AI deployment continued to increase, driving the cloud business towards a higher-quality structural upgrade.
Overall, AI-driven businesses generated over 11 billion yuan in revenue in Q4 2025, accounting for 43% of Baidu's core revenue, becoming a significant growth pillar for the company. The company launched Wenxin Yiyan 5.0 and completed restructuring of its model team, adhering to an application-driven approach and continuously optimizing multi-modal understanding, creative writing, and instruction-following scenarios. AI-native marketing services maintained high growth, generating 2.7 billion yuan in Q4 2025, a year-over-year increase of 110%. Digital humans and intelligent agents improved advertising efficiency and conversion rates, with AI applications generating over 10 billion yuan in annual revenue.
Conclusion
From the results,Baidu’s accumulation in cloud business has begun to translate into real growth momentum, with the Hongzhi Operator based on OpenClaw being the best proof.。
On the surface, it is just an AI application that allows users to 'raise lobsters,' but behind it lies a concentrated representation of a complete set of cloud capabilities, including a virtual mobile phone environment in the cloud, large model inference services, AI task scheduling systems, and security isolation mechanisms.
Baidu's ability to integrate such a complex technical system and ultimately deliver it to ordinary users in the form of 'download an app and start using' essentially reflects the company’s mature AI cloud infrastructure capabilities. This is also the direct reason why Peter Steinberger, founder of OpenClaw, posted two replies on overseas social platforms, calling China’s AI innovation speed 'Amazing.'
What surprised Peter Steinberger about Baidu will not stop here; it will gradually extend to a complete system of 'model + cloud + application ecosystem.' Whoever can truly connect computing power, platform, and applications is more likely to gain a long-term advantage in the AI era.
From this perspective,Red Finger Operator is not just a product; it is more like a concentrated showcase of Baidu's AI cloud capabilities.As more AI begins to perform tasks on behalf of humans, what truly determines the user experience is often not just the model itself but the invisiblecloud infrastructurebehind it. This is precisely Baidu’s earliest and most deeply cultivated trump card.
© THE END
All materials sourced from official public information
This article does not constitute any investment advice.
Risk Disclaimer: The above content only represents the author's view. It does not represent any position or investment advice of Futu. Futu makes no representation or warranty.Read more
Comments
to post a comment
1
2