OpenClaw puts an agent on your machine and a security hole free of charge

For weeks I've been seeing OpenClaw everywhere, sold as the free artificial-intelligence agent that finally runs on your own computer. KDnuggets introduced it as the free tool that's gone viral in 2026; DigitalOcean published its "what is OpenClaw" guide; and the project went from zero to more than a hundred thousand stars on GitHub in a matter of months; by late September it was at around 390,000. The argument that hooks you is privacy: no cloud, everything local.

What it does, specifically, is no small thing. OpenClaw reads and writes your files, runs commands on your system, browses the web and even sends emails on your behalf. It's an agent with its hands inside the machine, not a chat in a tab.

And that's where the news worth not skipping shows up. In February a critical remote-code-execution flaw in OpenClaw (CVE-2026-25253) was made public and, worse still, more than 135,000 instances were counted exposed on the internet, 63% of them with no authentication whatsoever, according to SecurityScorecard. Put another way: thousands of machines left the door open for a third party to do to them the same things the owner could. "Runs local" doesn't mean "is safe." And version 2.0, from late August, doesn't settle the matter: it adds shared cloud sessions, and its own documentation acknowledges that secrets in the shared store are kept unencrypted.

I think running processes locally matters, and that's why I'm interested in what OpenClaw proposes, even though the models you can run at home are still a long way from the commercial ones. That gap, what's more, is closing fast: in a year the conversation will be a different one.

But there's a question I can't shake. Are they really a hundred percent local and silent, or do they drag along some thread of a connection to the outside? I haven't seen any news of the latter, and I say it plainly so as not to accuse without proof. Even so, caution. If you like agents, run them on a machine that won't compromise your data.

Sources: KDnuggets · DigitalOcean · CVEFind (SecurityScorecard data) · InfoQ (OpenClaw 2.0) · OpenClaw documentation · gradually.ai