Security
Found something? Report it privately, and thank you — it is genuinely appreciated.
Reporting a vulnerability
Use the feedback form and pick the topic “A security problem”.
That form is a private channel. It does not post anything on the web: submissions go straight into a private inbox that holds nothing but reports, and it is closed — only the maintainer can read it. There is no public repository, no public issue list, and no other place a report could surface before a fix exists. Nothing you write in that form is published anywhere. (The gallery’s form is the one that publishes, after review — do not send a vulnerability there.)
Leave a way to reach you if you want a reply — an email address, or a Discord, LinkedIn or GitHub handle; the form asks which. The field is optional, and without it a report is genuinely anonymous, which also means there is no way to come back to you with questions or to credit you.
A useful report says what you did, what happened, and why it matters. A proof of concept helps enormously. Report it even if you are not sure it is exploitable — deciding that is the maintainer’s job, not yours.
One limitation, stated plainly: the form is protected by HTTPS in transit, but there is no end-to-end encrypted intake and no PGP key. If a finding is severe enough that you want an encrypted channel before describing it, send a short note with no details and ask for one.
What to expect
- Acknowledgement within a few days. This is a small project run by one person, not a company with an on-call rota — but reports are read and answered.
- An honest assessment. If it is a real issue you will be told so, along with the intended fix and a rough timeline. If it is not, you will be told why.
- Credit, unless you would rather not have it. The release notes will name you if you want to be named, and leave you out if you do not.
- Coordinated disclosure. Details go into the release notes once a fixed release is out. If a fix is taking a long time, that is a conversation, not a wall of silence.
Safe harbour
Research conducted in good faith under this policy is welcome, and no legal action will be pursued over it. That means: test against your own installation, do not access other people’s data, do not degrade the service for anyone else, and give a reasonable chance to fix before publishing.
Out of scope: volumetric denial of service (flooding a host proves nothing that is not already known), social engineering, physical attacks, spam, and reports generated by a scanner with no analysis behind them. Findings against this website and against the app are both in scope.
In scope, and wanted: a small, well-formed request to the hosted render door that costs a disproportionate amount of work. Routing is superlinear, and an algorithmic-complexity finding there is the risk that matters most. Send the input, not the flood.
What the design already rules out
Worth knowing before you dig, because it removes whole categories of bug from the equation:
- There is no account, no session and no cookie — so no authentication bypass, no session fixation, no CSRF against a logged-in user.
-
There is no database, and every page is a static file. Four things
accept input. Two are forms: feedback and the gallery’s. One is the
hosted render door —
/api/*andPOST /mcp, in open beta — which does render what you send it, in a sandboxed headless browser inside a hardened container, and then forgets it: your document is held for one request and never written to a log, a database or a file. That door caps how big a document may be, how many elements it may contain and how long one render may take, and it renders one document at a time per container; it has no per-caller quota during the open beta, and a finding against it is very much in scope. - The fourth is the collaboration relay, which carries a shared session’s traffic between the people in it. It cannot read what it carries — collaboration is end-to-end encrypted — and it takes a connection from anyone, because the people in a session are not accounts. A way to make it read, keep or misroute that traffic, to reach a session you were not invited to, or to tie it up for everyone else with a few small requests is in scope; flooding it is not.
-
The editor sends nothing about you or your diagrams. They stay on your
machine; the editor itself has no upload path to attack. Two places on this site take a
file, and only one you pick by hand. The feedback form takes a
.zip: it is bounded in size, never unpacked or inspected on the way through, and stored under a name this site generates — nothing from the uploader reaches a path. The gallery’s form takes one screenshot: a PNG, JPEG or WebP recognised by its bytes rather than its name (an SVG is refused), at most 4 MB, read no further than its header on the way through, and filed under a name this site generates in a private repository, where it waits for review; what is published is a re-encoded copy, made by the maintainer, never the uploaded file. Both are rate-limited, and the gallery stops taking submissions while too many wait for review. The one thing the editor asks for is itself: now and then it checks this site for a newer version of its own code, which is what puts the “Update available” prompt on screen — same origin, no third party, no identifier — and the new version is only applied when you click Reload. The desktop build checks this site’s release feed the same way, and that check is off until you turn it on — and when a check you allowed finds a newer version, the app downloads that release from this site, checks it against its published checksum and the developer’s signature, and installs it when you quit, unless you turn that off too. See privacy, which describes it in full. -
The pages ship a strict Content-Security-Policy with no
unsafe-inlineand nounsafe-eval, and load nothing from a third-party origin — with one exception, the Ko-fi tip panel on donate, which is not requested at all until you press the button.
None of that makes the project immune, and a finding that contradicts any of the above is exactly the kind of thing worth reporting.
What is inside each release
Every release from 0.110.0 publishes a software bill of materials beside its downloads, so that when a vulnerability is announced in a component you can see for yourself whether a Drawbridge build carries it. There is one for the desktop app, which also covers the web editor it embeds and this site hosts, and one for each self-hosting Docker image. They are CycloneDX JSON — the Electron, Chromium and Node.js the app runs on, every library built into it at its exact version with its registry hash and licence, the connector router at its pinned source commit — and they are generated by the release itself from the same build the release audits, never written by hand.
Each bill is signed with the release key, the same Ed25519 key the desktop app checks
before it installs an update of itself. The public half is
release-signing-key.pem,
and OpenSSL verifies a bill against it with no tool of ours; the command is on the
download page, beside the files. The key never leaves the
machine that makes releases: not the server that serves the downloads, not the build
machines, not the repository.
The bills describe the product. What runs this website — the gateway, the relay, the containers’ base images — is deliberately not published: it is our infrastructure, not your software, and its exact versions are the one place where publication would help an attacker more than a reader. Those versions are kept current under the same rule as everything else here (a security release is a task the day it is seen) and are checked by the release gate; a finding against them is in scope.
security.txt
This policy is advertised in machine-readable form at /.well-known/security.txt, per RFC 9116.