Is OpenClaw safe? Permissions, approvals and your data
OpenClaw can be run safely, but only if you do the work, because it hands the agent your machine by default. Tools execute directly on the host unless you add a sandbox, and the project's record includes 640+ GitHub security advisories.
Checked October 20, 2026. We summarise what OpenClaw Foundation says and flag what it doesn't say. We don't speculate.
The consent form
Five things to know before you connect your accounts. Each line is from OpenClaw Foundation's own pages unless it says otherwise.
- Where it runs
- On your own machine or server; the gateway binds to loopback on a normal host install, while container images default to an exposed bind and need auth.
- What it can reach
- Tools run on the host for the main session unless you configure a sandbox (Docker, Podman, SSH and others are documented).
- When it stops to ask you
- Unknown DM senders get a pairing code you must approve (openclaw pairing approve); groups are allowlisted and usually mention-gated. Exec approvals are documented on a separate page we did not read.
- What happens to your data
- Runs on your hardware; data goes to whichever model and tool providers you configure.
- How to cut it off
- Not described on the security overview page; an 'openclaw security audit' command checks configuration drift.
How to use it carefully
Everything runs on hardware you own. A normal host install binds the gateway to loopback, so it is not reachable from outside, while container images default to an exposed bind and need authentication first. Without a configured sandbox, the main session's tools run on the host itself. Unknown DM senders receive an 8-character pairing code that expires after an hour, and groups are allowlisted and usually need a mention. Messaging agents can send across conversations and providers unless you restrict that. Your data goes to whatever model and tool providers you configure. OpenClaw assumes one trust boundary per gateway, so its docs recommend separate gateways and credentials for people who do not trust each other. The 'openclaw security audit' command flags drift from safe defaults. Revoke steps were not on the pages we read.
Sources
- Unknown senders on DM-capable channels are paired by default and must be approved by the owner. github.com official, read Oct 20, 2026
- Tools run directly on the host for the main session unless sandboxing is configured. github.com official, read Oct 20, 2026
- The security overview says agents with the message tool can send across conversations and providers by default unless restricted. docs.openclaw.ai official, read Oct 20, 2026
- The GitHub security advisories page lists roughly 640+ published advisories (65 pages); the newest shown, dated 30 June 2026, include several rated High for authorization bypasses. github.com official, read Oct 20, 2026
- In February 2026 Koi Security reported 341 malicious skills on ClawHub, most delivering the Atomic Stealer macOS malware via fake prerequisites. thehackernews.com news, read Oct 20, 2026