IncidentAgent behavior
OpenClaw: one-click takeover of a personal AI agent
HighOpenClaw
OpenClaw, a popular self-hosted personal AI agent that reads messages, runs commands and holds API keys, could be taken over with a single malicious link. Researchers also found tens of thousands of OpenClaw instances exposed on the internet.
Replay the attack step by step in the Learning Center
How it happened
- The victim clicks a crafted link or visits an attacker's page.
- The page opens the OpenClaw control UI with a gatewayUrl pointing at the attacker, and the UI connects automatically, sending its stored token.
- With the token, the attacker connects to the victim's local gateway through the victim's own browser.
- The attacker turns off approval prompts, moves tool execution out of the sandbox and runs any command on the machine.
Impact
Full takeover of the user's computer and every account the agent could reach, even when the gateway listened only on localhost.
The fix
Fixed in OpenClaw 2026.1.29, which asks the user to confirm a new gateway URL before connecting.
Lessons
- Localhost is not a security boundary when a browser can reach it.
- Safety controls such as approval prompts should not be switchable by anyone who merely holds a token.
- Inventory and patch self-hosted agents, and keep their control panels off the internet.