Skip to content

Security: sigrix-io/gatehouse

Security

SECURITY.md

Security Policy

Gatehouse renders output produced by an AI agent, in a browser, on a page that belongs to somebody else. That is the whole of its attack surface and it is worth naming plainly: every string this package puts on a page came from somewhere the host does not control.

Reporting a vulnerability

Do not open a public issue.

Email security@sigrix.io with enough detail to reproduce — ideally the describe, status or stream bytes that trigger it. You will get an acknowledgement within three working days. CONTRIBUTING.md publishes the queues everything else waits in; security reports jump all of them.

What is in scope

  • Anything that gets agent output executed or interpreted as markup. innerHTML appears nowhere in this package and every agent string reaches the page through textContent; a path that escapes that is the most valuable report we can receive. Markdown rendering is block structure only, and inline markers are left as typed because an anchor whose href comes from agent output is exactly the hole the ban exists to close.
  • A second origin. Every request is supposed to go to the base URL the caller passed and to the protocol's path prefix. Anything reaching another origin — a font, an icon, a telemetry call — is a finding even if it looks harmless, because a page embedding this client did not agree to it.
  • A run started that the user did not ask for, or started twice. A reopened stream is a new run and spends real money (SPEC §4.5).
  • Anything a hostile describe or composition document can do. Both are authored by a seller and can come from any distributor; this package re-checks what it is handed rather than trusting it. A document that gets past that is in scope.
  • Leaking a credential into the page or a request. Postern carries none by design (§4.1.3); a path here that carries one anyway is a finding.

What is not in scope

  • A runner that is wrong. Postern places obligations on runners, and one breaking them is that runner's defect. If a runner's breach makes this package do something unsafe, that is in scope here — say so.
  • CORS and Private Network Access. Whether a page may reach a runner at all is the browser's decision and the runner's configuration (§2.3, §7). This package cannot widen it and must not try.
  • Weaknesses in the Postern specification itself. Report those to sigrix-io/postern — but if the specification is what led us here, say so and we will treat it as one.
  • A host that styles or embeds this badly. Rendering it inside a page that already has an XSS is not a finding against this package.

Supported versions

Version Status
0.1.x Pre-release — supported

Pre-1.0, and VERSIONING.md is blunt about what that means. There is no back-porting because there is nothing to back-port to.

What happens next

We will agree a disclosure timeline with you rather than impose one, and the changelog entry will credit you unless you would rather it did not.

There aren't any published security advisories