Security Policy
Thank you for helping to keep the internet secure. Please read this before making a security report.
Supported Versions
We support the last two major versions and their latest minor.
For non-security bugs, see thePuma release support policy.
| Version | Supported |
|---|---|
| Latest release in 8.x | ✅ |
| Latest release in 7.x | ✅ |
| All other releases | ❌ |
Our Threat Model
This determines what becomes a security advisory in Puma. We may make separate hardening fixes for things outside of this model.
Who do we defend against?
- Anyone who can reach the app port (the Unix, TCP or SSL bind that
Pumaserves theRackapp over). They can send any bytes, split up their data however they like, at any rate, with as many requests as they like on one connection or many. Their data is untrusted after it reaches your app.Pumamust hold up its side of theRackspec: build theRackenvaccurately, callcloseon every response body, and keep attacker controlled values in your return value ofapp.callfrom breaking the HTTP response. If an app that relies on theRackspec leaks data between requests or users becausePumadidn't follow it, that's in scope. - Anyone who can watch or change encrypted network traffic between the
pumactlutility and aPumacontrol server. - Anyone who can reach Puma's control server without the control token.
Who do we trust?
We don't defend against someone who already controls:
- Puma's config, command-line flags, environment variables, Gemfile, or load path
- Any code running inside
Puma, including your app,Rackmiddleware, plugins, and hooks - Direct calls to Puma's Ruby code, such as building a parser by hand
- Local files and process setup: state files, PID files, signals, sockets passed in by systemd or a parent process,
key_password_command, andpuma-wild - The control token
- A user account on the server (the actual host, not the
Rackapp). That account can reach Unix-socket control servers, state files, and logs. - Your reverse proxies and load balancers, when configured correctly. We don't trust their forwarded requests.
- If you tell
Pumato trust a header or the PROXY protocol, only your proxies should be able to reachPuma.
Bugs
Our features should work as designed. Bugs in our security features count in our threat model, even if they are optional. For example, if you turn on the optional control server and pumactl doesn't check a TLS certificate when it's supposed to, that's in scope.
What's a Denial of Service?
If you throw enough traffic at any web server, it eventually will fall over. So what's a DoS vulnerability in Puma?
- A small effort in
Pumacan cause outsized harm toPumaor other users, such as a process crash or restart. - The attack gets around a limit enforced by Puma, such as timeout on how long a connection can be held.
Slow clients
Puma writes each response from one of its request threads. A client that reads a large response slowly keeps that thread busy until the response is sent, so a few slow readers can occupy every thread. This is a known limitation #3110), not a vulnerability. If untrusted clients can reach Puma, put it behind a reverse proxy that buffers responses.
A report about slow clients is in scope when it gets around a limit Puma does enforce, such as a client that stops reading and holds a thread for longer than the write timeout.
Logging
Where your logs go (stdout, stderr, files, pipes, log shippers) is part of your process setup, which we trust. Puma writes logs synchronously and on a best-effort basis. If the log destination fills up, returns errors, or stops reading, Puma may slow down or stop serving requests. While the log destination is healthy, logging must not crash Puma.
A DoS attack that uses logging must still meet the "outsized harm" rule in "What's a Denial of Service?" above.
Puma shouldn't deliberately log secrets such as the control token or private keys.
Request smuggling
When Puma and a proxy disagree about where one request ends and another begins, you can get a smuggling vuln.
A smuggling report to Puma needs all three of these:
- A real, popular, maintained proxy, set up in a realistic or valid way. Most reports to
Pumafail this criterion because they propose hypothetical proxies which do not exist or would be doing something stupid. In that case, we might hardenPumaagainst it, but it's not going to get an advisory from us. - A reproduction against that proxy.
- Harm that crosses a security boundary, such as another user, cache, or connection shared with other users.
If Puma misreads an attacker's requests only for the attacker's own connection, that's not smuggling. We may fix it as a parser bug.
How we triage
We ask:
- Is it within the scope of our threat model?
- What's the CVSS 4.0 score?
The answers can lead to the following outcomes:
- Out of scope. It's a real bug, but outside our threat model (e.g., an attacker who has to be able to read a state file). May still get a hardening fix.
- CVSS score is 0.0. No harm shown. No advisory. May get a hardening fix.
- Advisory. We fix and release with a GHSA or CVE.
We don't publish advisories for unreleased code on main.
Reporting a Vulnerability
AI reports are welcome.
Rejected out-scope reports can be filed as public issues or PRs, but we will probably do that for you.
You'll help us most if you include Puma, Ruby, and OS versions, the config, a runnable reproduction, and, for smuggling, the proxy name, version and config.