Security

Your machine is the line

DorkOS is one place for every AI agent you run on your own computer. The one idea that explains everything else: your machine is the trust boundary. By default there is no account and DorkOS answers only to your own computer, the same way other developer tools do. It does not try to protect you from other software you already chose to run under your own user. It protects the line between your machine and everyone else.

Reaching DorkOS from elsewhere requires a login

The moment you want to reach DorkOS remotely, by starting a tunnel or opening it to your network, it becomes a networked service. So DorkOS requires an owner account before it will do that. Start a tunnel with login off and it walks you through creating one first. Bind to a public network address with no login and it refuses to start. Remote visitors get a sign-in screen, and every request needs a valid session.

Running an agent is a real act of trust

Agents can read and write files and run commands inside the boundary you give them. Installing a package from the Marketplace runs that package’s code on your machine. We fetch packages over safe channels and roll back a failed install cleanly, but we do not yet verify a package signature for you. Install packages the way you would run any script from the internet: only from authors you trust.

Four things an agent cannot do without asking you

Your agents can operate DorkOS itself, not just write code. Most of what they do is easy to take back, so it just happens and lands in your activity feed. Four things are not easy to take back, and DorkOS stops and asks first: removing an installed package, deleting a scheduled task, removing an agent, and changing an agent’s safety boundaries. Installing a package or creating one asks too, because both bring new code onto your machine. You get a card describing exactly what would run, written by DorkOS rather than by the agent, and an agent that identifies itself is refused when it tries to answer its own card.

Here is the limit, because it matters more than the feature. With no login, which is how DorkOS starts, DorkOS cannot tell you apart from other software running as you on the same computer. It refuses the agent that asked, and it records who answered, so nothing happens quietly. But a program on your machine that deliberately hides what it is looks like you. Turning on login is what closes that, and then only your signed-in account can answer a card. We would rather write that down than let the word “approval” do work it cannot do.

Your AI keys, your data

You bring your own AI keys. Your prompts and code go straight to the model vendor you chose, under their terms. DorkOS passes it along and keeps no copy. See our Privacy Policy for the full picture.

Reporting a vulnerability

If you think you have found a security issue, please tell us privately first. Email security@dorkos.ai or use the private “Report a vulnerability” button on our GitHub Security page. We are a tiny team, so we aim to reply within three business days and fix confirmed issues before moving on to new features. We will credit you if you would like.

The full policy lives in SECURITY.md, and the detailed threat model explains what DorkOS trusts and where the sharp edges are.

When this changes

DorkOS moves fast. When how it handles security changes, we update this page and change the date at the top. No quiet edits.