Skip to content

Security: pt-act/ultimate

Security

SECURITY.md

Security policy

Reporting a vulnerability

Report through GitHub's private vulnerability reporting: Security → Report a vulnerability on this repository. Do not open a public issue for a suspected vulnerability.

Include what you know: the affected area (server/request-security.ts, a specific extension, the web UI), the commit you tested, a reproduction and, if you have one, an impact sketch. A proof that needs no special access is worth more than a long theory.

Reports are triaged as they land. You get an answer before anything is published; fixes land on main like any other change, and credit is yours to claim or decline.

Scope

Ultimate runs on your machine: the server binds loopback only, and the UI is meant for the operator using it. The optional tailscale extension serves it over the tailnet, to the logins it allows. In scope is anything that lets a web page, a remote server or a repository break that boundary —

  • a request admitted that server/request-security.ts should refuse: a cross-site write, a cross-origin read, a frame the header policy does not name;
  • the agent's tools or extensions reading or sending data past a limit their code enforces;
  • a page in the UI acting as another conversation or project (?conversation=, previews, window.ultimate);
  • secret material (provider keys, tokens, conversation contents) ending up somewhere the operator did not put it;
  • a tailnet request admitted for a login the tailscale allowlist does not name;
  • a project's .ultimate/ extensions, agents, skills or settings taking effect before the operator trusts that project.

Out of scope:

  • a hostile process already running on the machine, or an extension the operator chose to install;
  • the browser sandbox disabled with ULTIMATE_CHROMIUM_SANDBOX=0 — the knob exists for environments where sandboxing cannot start, and turning it off there is the operator's call;
  • attacks that require the operator to hand over access: a prompt they paste, a link they accept;
  • content an agent reads (a web page, a file, a tool result) steering what it does, unless it gets past a boundary enforced in code: project trust, the request and header policy, the Tailscale allowlist.

Supported

Security fixes land on main; run the current main to get them.

There aren't any published security advisories