wisp runs coding agents on your code, and wisp Cloud runs them on our server, so we take reports seriously.
Report it privately, not in a public issue:
- GitHub: Report a vulnerability (private advisory).
- Email: hello@trywisp.dev, with "security" in the subject.
Include what you found, how to reproduce it, and the version (wisp --version) or, for wisp Cloud,
the time it happened. We reply within three working days, keep you posted while we fix it, and
credit you in the release notes if you'd like.
- wisp itself: the local server, the page, the CLI, the desktop app and the Android app.
- wisp Grid and the hosted wisp Cloud at
grid.trywisp.dev: sign-in, the edge, the tunnel and the isolation between users' boxes. - The install script at
https://trywisp.dev/install.shand the release downloads.
Spells are meant to be sandboxed (no network, no page, no wisp API): a way out of that sandbox is in scope too.
wisp keeps an agent to its own task: with its own token it can't raise a task's mode, and a change to your setup it asks for (a connector, a crew member, a project) is a card you allow or deny; nothing runs before you allow it, and the agent can't allow it itself (ADR-0112, ADR-0113).
That isn't a wall against an agent that can run commands on your computer as you. wisp's server treats a request on
its own port with no token as you, and an agent with a shell can also edit your files or run wisp itself. So:
- In Plan, and in Edits where each command waits for your OK (Claude Code), the agent has no shell of its own.
- In Full auto, the agent runs commands as you: give it work you'd trust it with.
- For work you don't trust, use a sandbox (
/sandbox, ADR-0047): it runs on a machine of its own, and what it can reach without a token is that sandbox's wisp, not your computer's.
A report that an agent got past these without a shell (or out of a sandbox) is in scope.
Fixes land in the latest release. Update with npm install -g wisp-agents@latest, or let the app
update itself.