OpenClaw emerged in late 2025 as one of the fastest-growing open-source projects in GitHub history, promising to transform AI from a conversational tool into an autonomous “digital employee”. Unlike a standard chatbot, OpenClaw can clear an inbox, manage a calendar, execute shell commands, and browse the web on a user’s behalf all through the messaging apps people already use.
But that power comes with a documented security crisis: hundreds of vulnerabilities, a large-scale supply-chain attack, and over 100,000 instances exposed to the public internet. This guide explains what OpenClaw actually is, how it works, its genuine use cases, and the risks that every individual and organization must understand before adoption.
Key Takeaways
- OpenClaw is a self-hosted gateway, not a cloud service. It runs on your hardware, connects chat apps to AI models, and executes tasks locally with the permissions of your user account.
- It performs real actions, not just text generation: file management, email triage, calendar scheduling, shell commands, and browser control.
- Its security maturity lags far behind its adoption. As of March 2026, the platform had accumulated 169 security advisories (five critical), and researchers identified over 450 vulnerabilities in the framework.
- Enterprise adoption is widespread but risky. Roughly 22% of surveyed enterprises reported shadow-IT use within months of launch, while over 135,000 instances were exposed to the public internet.
- Zero Trust governance is essential for any organization deploying OpenClaw. The Cloud Security Alliance has published a hardening framework specifically for this platform.
What Is OpenClaw? Definition and Architecture
OpenClaw is a free, open-source, self-hosted gateway that bridges messaging applications WhatsApp, Telegram, Discord, iMessage, Slack, and others to AI coding agents and large language models. It was launched in November 2025 by Austrian developer Peter Steinberger under the name Clawdbot, renamed Moltbot in January 2026, and finally settled as OpenClaw shortly thereafter.
The core architecture consists of four components: a Gateway (the single source of truth for sessions and routing), an Agent (the reasoning layer that connects to external LLMs), Skills (extensible capabilities accessed through the ClawHub marketplace), and Memory (persistent context from past conversations and actions). Crucially, while the reasoning layer may run in the cloud, execution occurs locally on the user’s machine, meaning the agent inherits the same permissions as the user’s account.
This design is what separates OpenClaw from ChatGPT or Claude. A chatbot responds; OpenClaw acts. It can execute shell commands, read files, control a browser, and operate without human oversight for tasks such as managing files or writing code.
Core Features and Capabilities
OpenClaw’s feature set is defined by its tool-using agent framework and its multi-channel accessibility.
| Capability | What It Does |
| Local system access | Executes shell commands, reads/writes files, and controls the local operating system |
| Browser automation | Navigates websites, fills forms, and extracts data without manual intervention |
| Email and calendar management | Triages inboxes, drafts and sends messages, schedules meetings, and checks users in for flights |
| Multi-channel gateway | Operates through WhatsApp, Telegram, Discord, iMessage, Slack, Microsoft Teams, Signal, and more |
| Persistent memory | Remembers previous conversations, past data, and actions to improve task continuity |
| ClawHub skills marketplace | Extends functionality through community-built skills and integrations |
| MCP integration | Connects to thousands of third-party applications via the Model Context Protocol |
How OpenClaw Is Used: Practical Scenarios
Personal Productivity
For individual power users, OpenClaw functions as an always-available personal assistant. A user can message it through Telegram to clear an inbox, summarize a document, draft a reply, or schedule a meeting. Because it runs on local hardware, sensitive data does not need to be routed through a third-party cloud service for execution.
Enterprise Workflow Automation
In business settings, OpenClaw can be integrated into existing workflows through APIs, scripts, and MCP connections. Documented use cases include triaging email, running reports, interfacing with internal tools, and automating development tasks. The platform’s self-hosted nature appeals to organizations that want to retain control over data while still leveraging frontier AI models.
The “Digital Employee” Narrative
Industry leaders at Nvidia’s GTC conference in March 2026 described OpenClaw as a landmark shift moving AI from answering questions to taking real-world action. Nvidia CEO Jensen Huang said the platform is “bringing agentic systems to the consumer mindset”. In China, users coined the phrase “raising a lobster” (the platform’s icon is a red lobster) to describe training the assistant for increasingly complex tasks.
OpenClaw Voice, Media, Canvas, and Mobile Features

OpenClaw supports more than text-based AI automation. Depending on the connected messaging platform, installed skills, and device configuration, users can access voice interactions, exchange media files, and use visual interfaces for supported tasks.
- Voice Features: Supported integrations can process voice messages and convert speech into text for AI-assisted tasks.
- Media Support: OpenClaw can handle supported images, audio files, and documents through compatible messaging channels.
- Canvas Features: Compatible clients can display interactive visual content, dashboards, and task-related information.
- Mobile Access: Users can interact with OpenClaw through messaging apps on Android and iOS without operating the gateway directly from their phones.
These features extend OpenClaw’s usefulness for personal productivity and business automation, although availability depends on the platform, integration, and configuration.
The Security and Governance Crisis
The same capabilities that make OpenClaw powerful also make it dangerous when deployed without safeguards. The documented risk landscape is extensive.
Vulnerability Scale
As of March 2026, OpenClaw had accumulated 169 security advisories five rated Critical, 58 High, 96 Medium, and 10 Low. A separate Kaspersky audit identified 512 vulnerabilities, eight of them critical.
By early May, researchers had reported at least 454 vulnerabilities in the framework through the National Vulnerability Database. The flaws include remote code execution, Docker sandbox escapes, path traversal, and access-control errors.
The ClawHavoc Supply-Chain Attack
A campaign called ClawHavoc planted between 341 and 1,467 malicious skills on ClawHub, the community marketplace. The attack combined prompt injection with the Atomic macOS Stealer to harvest credentials across thousands of deployments. This exploited a systemic weakness in OpenClaw’s third-party skills ecosystem.
Internet Exposure at Scale
SecurityScorecard’s STRIKE team identified over 135,000 OpenClaw instances exposed to the public internet across 82 countries. Other scans placed the figure near 180,000. Many of these deployments lack basic authentication or network segmentation, creating a large attack surface.
Architectural Flaws That Patches Cannot Fully Close
In June 2026, two independent research teams demonstrated a trusted-input flattening vulnerability: OpenClaw agents execute attacker-controlled instructions embedded in objects treated as trusted. This is an architectural flaw that “no patch can fully close,” requiring organizations to treat all external data objects as untrusted inputs.
The “Devil’s Bargain”
CrowdStrike CTO Elia Zaitsev described AI agents like OpenClaw as “a devil’s bargain”: the greater the power and the more use cases, the greater the risks introduced. Chinese authorities issued multiple alerts warning that OpenClaw could expose organizations and individuals to significant cybersecurity vulnerabilities.
Enterprise Adoption and Zero Trust Governance
Despite these risks, OpenClaw spread rapidly through enterprises. Within months of its January 2026 launch, it achieved 22% shadow-IT penetration among surveyed organizations, accumulating over 250,000 GitHub stars. This pattern employees adopting a powerful tool faster than IT can govern it creates a familiar but acute governance challenge.
The Cloud Security Alliance has published a comprehensive Zero Trust hardening framework for OpenClaw, structured around several core requirements:
- Agentic control plane: Separate security policy from agent execution, with identity governance, authorization enforcement, and behavioral observability.
- Skill vetting and runtime sandboxing: Treat the ClawHub marketplace as untrusted; review and isolate third-party skills before deployment.
- Credential and file-system management: Limit the permissions granted to the agent; do not allow it to inherit full user-account privileges by default.
- Behavioral monitoring and incident response: Implement playbooks for detecting anomalous agent behavior, including unexpected API calls or file access.
Nvidia responded to the security concerns at GTC by unveiling Nemo-Claw, an open-source stack designed to layer privacy and security controls onto OpenClaw. The project itself transitioned to an independent 501(c)(3) foundation with OpenAI’s support in February 2026, a structural change intended to support long-term governance.
Conclusion: What Is OpenClaw
OpenClaw represents a genuine leap in what consumer and enterprise AI can do. It moves AI from conversation to action, and its open-source, self-hosted design gives users control that hosted services cannot match. But that control is only as safe as the governance around it. The documented record of hundreds of vulnerabilities, a large-scale supply-chain attack, and over 100,000 exposed instances shows that adoption has far outpaced security maturity.
For individual users, the guidance is straightforward: run OpenClaw only on isolated hardware with minimal permissions, avoid untrusted ClawHub skills, and treat every external input as potentially hostile. For enterprises, the path forward is not abandonment but disciplined adoption: a Zero Trust architecture, a dedicated agentic control plane, and a clear policy for shadow-IT containment. The platform’s future depends less on its capabilities than on whether its community and corporate stewards can close the gap between what it can do and what it can safely be trusted to do.
What Is OpenClaw FAQs
1. What Is OpenClaw and How Does It Work?
OpenClaw is an open-source AI agent platform that connects messaging apps with AI models. It automates tasks such as managing emails, organizing files, browsing websites, and executing commands.
2. Is OpenClaw Free and Safe to Use in 2026?
OpenClaw is free and open-source, but AI model APIs and hosting may cost money. Security depends on proper configuration, restricted permissions, and avoiding untrusted third-party skills.
3. What Is OpenClaw Used for in Business?
OpenClaw helps businesses automate repetitive tasks, manage workflows, generate reports, and support development activities. Organizations should apply strict access controls and monitor agent actions to protect sensitive data.