OpenClaw

What Is OpenClaw?

OpenClaw is a self-hosted AI agent runtime that runs on your own hardware, using a large language model to execute real tasks autonomously. Rather than generating a response and stopping there, it can run shell commands, manage files, browse the web, handle email, and chain those actions together across multiple steps without requiring you to prompt each one. 

Launched in November 2025 under the name Clawdbot, the project surpassed 200,000 GitHub stars within three months, making it one of the fastest-growing open-source projects on record. Its creator, Austrian developer Peter Steinberger, subsequently joined OpenAI to lead their personal agents division, with the project moving to an independent open-source foundation.

How Does OpenClaw Execute Tasks?

OpenClaw runs as a single Node.js process on your machine, called the Gateway, which routes incoming requests to the appropriate agent session. When a task arrives, the agent assembles context from your conversation history and workspace files, sends that to your chosen language model, such as Anthropic’s Claude or OpenAI’s GPT, and executes any tool calls the model requests before returning a response. This loop repeats until the task is complete.

What Are The Security Risks of Running OpenClaw?

OpenClaw’s power and its risks share the same root: broad system access. An agent that can execute shell commands, read and write files, and process external content from emails or websites creates a significant attack surface.

Prompt injection is a primary threat, where malicious instructions embedded in a document or webpage the agent reads can redirect it toward unintended actions. To that end, security researchers have found over 800,000 exposed OpenClaw instances on the public internet as of May 2026, and threat actors have also regularly distributed malicious skills via the ClawHub registry.

For anyone running OpenClaw in environments where real value is at stake, software-level guardrails are insufficient on their own. Ledger’s 2026 AI Security Roadmap addresses security gaps like this directly: the human defines policies, a Hardware Security Module enforces those policies on every subsequent agent action, and anything outside the approved boundaries is rejected at the hardware level before execution. Agents operate autonomously inside those limits, with hardware as the enforcing layer rather than code.