OpenClaw: AI marvel or cybersecurity nightmare?
People have flocked to the OpenClaw artificial intelligence agent since it was launched in November by Austrian programmer Peter Steinberger. The digital assistant can use your computer to handle complex tasks that previously only a human could undertake, such as making travel bookings, prioritizing emails and drafting replies, surveying product catalogs and emailing vendors.This leap in productivity comes with a catch: OpenClaw has proved to be a gift to hackers. One critical flaw, dubbed ClawJacked, allowed intruders to take control of a user’s OpenClaw agent simply by getting them to visit a malicious website. That defect was fixed. But researchers have found more than 40,000 vulnerabilities in the software.Nowhere is there as much excitement or apprehension around OpenClaw as in China, where its rapid adoption has led to gyrations in the stock prices of big local tech firms and prompted officials to warn government agencies and state-owned enterprises — including some of the country’s largest banks — against installing it on office devices. What is OpenClaw?It’s an AI assistant that can be set up on a computer or even a smartphone. Giant AI companies including OpenAI Inc and Anthropic PBC also offer agents that field tasks for users. However, those companies don’t allow customers to modify their agents’ underlying parameters. OpenClaw’s code is “open source,” which allows users to be more freewheeling with the product, opening the door to some more inventive — and potentially risky — uses for the technology. OpenClaw works from the data on a user’s phone or computer, in contrast to other popular AI services that process it remotely in so-called cloud networks.Steinberger, an established figure in the Apple software ecosystem, released the app under the name Clawdbot, briefly renamed it Moltbot and finally settled on OpenClaw. The project coincided with a growing fascination in the tech world with AI agents, and quickly outgrew its creator to become a community-driven enterprise with thousands of contributors worldwide.Steinberger has since joined OpenAI to help the world’s most valuable private AI company build the next-generation of agents. OpenClaw remains a separate and independent entity under a foundation structure. How does OpenClaw work?OpenClaw works inside popular messaging platforms such as WhatsApp, Telegram, WeChat, Discord, Slack and Signal as an interface so users don’t need to get to grips with a whole new system.It requires some technical know-how to set up, but once that’s done, a user can instruct OpenClaw using natural-language commands via those apps. It recalls context and preferences across sessions and past interactions, tailoring responses and actions over time to better serve a specific user’s needs. It can execute commands on a computer, read files, install software and carry out multi-step tasks across apps including messaging platforms and productivity software.Anyone with specialized software skills can access OpenClaw’s underlying code and teach it to learn new “skills” so it can carry out an even wider range of functions. Software developers can also plug OpenClaw into other AI models to draw on their capabilities. The biggest of those offer their own powerful AI agents, but these are less easy to customize than OpenClaw. What tasks can OpenClaw handle?According to OpenClaw users, it can book a flight or a cab ride to the airport, schedule meetings, deal with an overflowing email inbox by prioritizing the most important messages, and read and act on messages independently if directed to do so. They say it can navigate websites, analyze files and documents such as PDFs, spreadsheets or batches of code and generate summaries or follow-up actions based on their contents. So-called power users say they’ve adapted the service so it’s adept in specific job roles such as personal procurement officer or product inventory manager. It can even negotiate prices with retailers while a user is offline. Why is OpenClaw so popular in China?Major Chinese cloud computing providers, including Tencent Holdings Ltd, Alibaba Group Holding Ltd and Baidu Inc, have rushed to offer OpenClaw to their customers for “one-click” deployment. And Chinese AI companies have jumped in to offer OpenClaw as a way to boost usage of their own platforms.Local authorities in tech hubs like Shenzhen, Wuxi and Hefei have announced subsidies of up to 2mn yuan (approximately $290,000) for OpenClaw-based projects and related hardware. Why are there concerns about OpenClaw?Some cybersecurity experts see a disaster in the making. In March, several Chinese government agencies and banks issued official alerts over OpenClaw. They detailed risks including data theft and “prompt injection” attacks, in which texts are sent to trick an AI agent to perform unauthorized actions.Hackers can create new “skills” for OpenClaw that include installing hidden malware and harvesting the personal data of users and their contacts. Kasimir Schulz, director of security research at HiddenLayer Inc, said OpenClaw ticks all the boxes when it comes to cybersecurity risk: It has access to private data, it can communicate externally and has exposure to untrusted content.For agentic AI like OpenClaw to be really useful, it needs to know all about you, and gain access to a variety of apps. That makes them juicy cyberattack avenues or targets.When a user’s system was compromised using the “ClawJacked” security vulnerability, attackers could read files, steal passwords, drain crypto wallets or leak sensitive chat histories.There are also anxieties around governance and accountability: Unlike proprietary AI platforms backed by government-imposed compliance frameworks, OpenClaw’s decentralized model leaves responsibility in the hands of individual users, with no central authority to enforce safeguards.The agent’s rapid adoption in China has sparked a debate there about regulatory oversight, with national authorities weighing the risks of a widely distributed AI agent that operates outside corporate or state-controlled ecosystems.OpenClaw’s creator, Steinberger, has told Bloomberg News the AI tool and its security are works in progress. “It’s simply not done yet — but we’re getting there,” he wrote in an email in February. “Given the massive interest and open nature and the many folks contributing, we’re making tons of progress on that front.”Steinberger said the main security breaches come from users not reading OpenClaw’s guidelines, though he acknowledged there was no “perfectly secure” setup. “The project is meant for tech savvy people that know what they are doing and understand the inherent risk nature of LLMs, he said. Does OpenClaw have rivals?The field of AI agents is ballooning. NanoClaw, Nanobot, and NullClaw are smaller apps with lower computer resource requirements. At the other end of the scale are the agents developed inside the big US AI ecosystems, where code can’t be accessed and tweaked by developers: OpenAI’s Operator, Anthropic’s Claude Code and Google’s Project Mariner.Then there are more specialized, business-oriented contenders such as Amazon’s AWS Bedrock Agents. And AI assistants are increasingly embedded in dedicated hardware devices such as Humane and Rabbit.