On March 8, according to CCTV News, the National Vulnerability Database (NVDB) under the Ministry of Industry and Information Technology issued an alert: it had detected that some instances of the open-source AI agent OpenClaw (nicknamed “the lobster”) pose relatively high security risks under default or improper configurations, and can easily trigger network attacks, information leaks, and other security problems.
“Raising Lobsters” Becomes a Buzzword, the AI Agent Breaks Out
Recently, this open-source AI agent — affectionately nicknamed “little lobster” by netizens because its icon is a red lobster — has gone viral across the internet. Multiple topics related to “raising lobsters” trended on Weibo, including “What on earth can an AI lobster do” and “Netizen makes 260,000 yuan in a few days installing lobsters for people at their homes.”
OpenClaw (formerly named Clawdbot, Moltbot) is an open-source AI agent software. By integrating multi-channel communication capabilities with large language models, it builds a customized AI assistant with persistent memory and proactive execution capabilities, supporting local private deployment.
For users, OpenClaw is like a round-the-clock “AI laborer” on standby. It can directly control the user’s computer, helping tidy the desktop, send emails, run self-media accounts, and automatically fix program bugs — as if it gave the computer a brain. Technically, when a user issues a complex instruction, the locally installed OpenClaw can break down the task, search the internet, call local software, recognize errors, self-correct and retry, send requests to a cloud large model’s API, and complete the task step by step.
Deployment Barriers Spawn an “Installer-For-Hire” Business, Enterprises Poach With High Salaries
Unlike zero-barrier products like DeepSeek and ChatGPT, OpenClaw’s local deployment requires fairly complex environment setup, which has spawned “installer-for-hire” services on social platforms. According to reports, remote installation starts at 300 yuan, and on-site service ranges from 500 to 1,500 yuan. One practitioner claimed to have earned 260,000 yuan in a few days off this craft.
On March 8, a Xianyu seller offering remote installation said: “Lately the demand is overwhelming; right now the queue after ordering is about 30 minutes.” The current charge is 150 yuan, and they expected it to rise to 200 yuan soon.
Corporate demand for OpenClaw talent is also surging. Job platforms show that multiple enterprises in Beijing, Shanghai, Shenzhen, and elsewhere have posted roles like “OpenClaw Development Engineer” and “AI Product Manager,” with monthly salaries ranging from 10,000 to 30,000 yuan, or roughly 120,000 to 360,000 yuan a year. The financial engineering teams of several brokerages — including Founder Securities, GF Securities, and Soochow Securities — have also released special reports teaching finance practitioners step by step how to deploy and use OpenClaw.
On March 4, Tencent Cloud posted an announcement on WeChat Channels, saying it would offer free OpenClaw installation at the North Plaza of the Tencent Building in Shenzhen, where queues formed outside the building.
A Security Risk Case: A User’s Emails Were Deleted Automatically and Couldn’t Be Stopped
OpenClaw’s total-control capability also brings significant security risks. According to The Paper, in February of this year Summer Yue, a researcher from Meta’s superintelligence team, shared her harrowing experience on X (formerly Twitter): her “lobster” suddenly started deleting emails in bulk, and she could barely stop it.
Summer Yue recalled that her instruction had been very simple: “Check this inbox and suggest emails you want to archive or delete.” She had also attached the usual “safe word” constraints, yet those constraints had no effect — OpenClaw went straight into mass-deleting emails. Until she resorted to the final method of a “physical shutdown,” there was no way to make the “lobster” stop.
MIIT Spells Out the Risk Points: Blurred Trust Boundaries, Privilege Escalation
In its alert, NVDB under the Ministry of Industry and Information Technology clearly pointed out that because OpenClaw has “blurred trust boundaries” when deployed, and has features such as continuous operation on its own, autonomous decision-making, and calling on system and external resources, in the absence of effective permission controls, auditing mechanisms, and security hardening, it may — due to instruction induction, configuration flaws, or being maliciously taken over — perform out-of-bounds operations, causing a series of security risks including information leaks and system compromise.
In response to these risks, MIIT advised relevant organizations and users, when deploying and applying OpenClaw, to:
- Fully check the public-network exposure, permission configuration, and credential management
- Close unnecessary public-network access
- Improve security mechanisms such as identity authentication, access control, data encryption, and security auditing
- Continuously follow official security advisories and hardening recommendations to guard against potential cybersecurity risks
Industry Views: AI Anxiety Drives the Craze, Ordinary Users Should Wait and See
Yuan Shuai, deputy director of the investment department at the China Urban Development Research Institute, told Hongxing Capital Bureau that behind OpenClaw’s viral rise is a tight connection to the AI anxiety pervasive today. As AI technology iterates rapidly, more and more people worry they’ll be made obsolete by AI and hope to actively master AI tools to turn them into a competitive edge — and OpenClaw happens to be an outlet that eases this anxiety.
Still, some AI industry practitioners believe that while OpenClaw is indeed a proactive AI assistant, unless the user is a programmer, the productivity boost is actually “not much”: “It’s not as smart as you imagine. How smart it is depends entirely on the large model you use, but good large models are very expensive.”
Zhou Hongyi, founder of 360, commented that OpenClaw has changed many people’s perception of agents — it’s the conceptual innovation of “raising an agent.” Kai-Fu Lee, who served as a global vice president at Microsoft and Google, said that future competition will happen at the level of multi-agent systems, which will become a bridge connecting the digital and physical worlds.
In Closing
As a developer who has long followed the development of AI technology, I think OpenClaw’s emergence does represent an important direction for AI applications — moving from passive response to proactive execution. However, the novelty of a technology is often accompanied by a blurring of security boundaries.
In the current AI craze, ordinary users need to stay rational even more: while enjoying the convenience of technology, be fully aware of the security risks that locally deployed AI agents may bring. After all, when a “lobster” can fully control your computer, it can be a capable assistant — or the source of a security risk.
For users who want to give it a try, my suggestions:
- Test in an isolated environment: deploy on a virtual machine or a dedicated device, and avoid using it directly in production
- Strictly control permissions: limit OpenClaw’s file-system and network access
- Stay vigilant: check the logs regularly and watch for abnormal behavior
- Follow security updates: keep up with official and community security advisories
Technological progress always comes with both opportunity and challenge; the key is how we find the balance between the two.
Parts of this article reference reports from CCTV News, The Paper, Hongxing Capital Bureau, and other media.
评论
0 条评论