Security disclosure

How to tell us about a security problem in instruction.md: where to send it, what to put in it, what is in scope, and what happens after you send it. Report security problems privately, through the form on Contact us. Don't report them in a public issue, a GitHub discussion or the specification's issues.

Report a problem

  1. Open Contact us

    Go to Contact us. The form is at the top of the page. It needs JavaScript: before it sends, your browser solves a small proof-of-work puzzle that tells the form a person sent it.

  2. Choose Security

    In What is it about?, choose Security.

  3. Fill in the form

    Field Required Limit
    Your name Yes 120 characters
    Your email Yes 254 characters. We answer at this address
    Organization (optional) No 120 characters
    Your message Yes 10 to 5,000 characters

    The form takes no attachments. Write the report in the message. If it doesn't fit, send a summary and say you have more; you can send the rest by replying to our confirmation email.

  4. Send it

    Press Send. The page shows your request's number, such as R-7K3M-Q9TD. Keep it. You also get an email, "We received your request R-…", with the same number. At most three confirmations a day go to one address, so a fourth request that day gets its number on the page only.

What to include

  • Where. The address, tool or package and its version: for example https://instruction.md/mcp and the tool's name, or instruction.md 0.11.0 on npm.
  • How to reproduce it. The steps, requests or code, from a fresh account where you can.
  • What an attacker gains. What they can read, change or do that they shouldn't, and what they need first, such as an account, a role in an organization, or a person clicking a link.
  • What you saw. The exact responses or error messages, with secrets removed.
  • Your plans. Whether and when you intend to publish, and how you would like to be named, if at all.

Never put a working credential in the report: not yours and not anyone else's. If your testing exposed someone else's data, say so, say what you saw, and delete your copy.

Scope

In scope:

  • The hosted service at https://instruction.md: the site, the editor, the MCP server at /mcp, the APIs under /api, and the sign-in pages under /auth.
  • The packages we publish: instruction.md, @instruction.md/agents, @instruction.md/mcp and @instruction.md/seal on npm, the render@1 script, the libraries for other languages and the VS Code extension.
  • The self-hosted stack as we ship it. A problem in an instance someone else runs that comes from their own configuration is theirs; tell its operator.
  • Sealed workspaces and their cryptography. They are beta, not yet independently reviewed, and we want to hear about weaknesses in them.
  • Signing and verification: a way to make a version look signed by someone who didn't sign it, or to make an agent accept a text it shouldn't.
  • The instruction format itself, where reading an instruction as The instruction.md Specification says leads to a security problem. Report it privately like any other problem, not in the specification's public issues.

Out of scope:

  • Services we don't run: OpenAI, which runs the editor's assistant; your organization's identity provider; the AI apps and agents you connect.
  • Denial of service by volume, and testing that degrades the service for others.
  • Social engineering of our staff or our users, and physical attacks.
  • Reports from automated scanners without a demonstrated effect, such as a missing header that enables no attack.
  • Problems that need a device or account that is already compromised.

Testing rules

  • Test only with accounts and organizations you own, or that their owners let you test.
  • Don't read, change or delete other people's data. If you reach some by accident, stop, and tell us in your report.
  • Don't degrade the service, and don't send the forms over and over: they are rate limited (below).
  • Don't publish details or tell anyone else until we have fixed the problem, or until we agree a date with you.

What happens next

  1. We confirm. Press Send. The page shows your request's number, and an email with the number goes to the address you gave. It is signed "instruction.md support", and a reply to it reaches our support mailbox.
  2. A person reads it. Every message is read by a person, not answered by a bot.
  3. We answer by email, usually within two working days. If we need more, we ask in that email.
  4. We keep you informed while we fix it, and tell you when it is fixed.

We don't promise a time to fix: it depends on the problem. We don't run a bug bounty, so we don't pay for reports.

Our email never carries a sign-in link or a code, and we never ask you for a recovery code. The confirmation says so too. An email that asks you for one of these isn't from us.

security.txt and PGP

There is no /.well-known/security.txt and no PGP key for reports yet, and we publish no security email address. The Contact us form, over HTTPS, is the one way to report. This section will name them when they exist.

If the form doesn't send

What the form says Why What to do
"Too many messages from here. Please try again in …" A rate limit: 5 messages an hour from one address, 20 a day from one network, 3 an hour and 10 a day from one email address, or the form's daily limit for everyone Wait the time it says, then send it again. Put several findings in one message
"We couldn't check this form was sent by a person. Please send it again." The proof-of-work puzzle failed or ran out (it lasts ten minutes) Send it again. The page fetches a new puzzle
"This form can only be sent from instruction.md. Open this page there and try again." The form was sent from another site or a copy of the page Open Contact us on instruction.md itself
"Something went wrong on our side and your message wasn't sent." A problem on our side Try again in a moment

If you can't sign in to your own account, that isn't a security report: see Lost your passkey? and Support.

Questions

How do I report a security vulnerability?

Through Contact us, choosing Security. Don't use public issues. A person answers by email, usually within two working days.

Is there a security email address or a PGP key?

Not yet. The form is the only way to report, and there is no security.txt yet. See security.txt and PGP.

Do you pay for reports?

No. We don't run a bug bounty. We do answer every report, and keep you informed while we fix it.

Can I test sealed workspaces?

Yes, with a workspace of your own. Sealed workspaces are beta, not yet independently reviewed, and reports about their cryptography are in scope. See Sealed instructions.

Last checked against the product:

Was this page helpful? Report a problem