diff --git a/SECURITY.md b/SECURITY.md index 965c5631..424e1484 100644 --- a/SECURITY.md +++ b/SECURITY.md @@ -34,11 +34,13 @@ Security issues are defects in Homebrew-maintained code or infrastructure that l Homebrew is designed around a single-user machine security boundary. Homebrew assumes the user running `brew` controls and trusts their own account, Homebrew prefix, cache, configuration, taps and shell environment. Shared multi-user Homebrew installations are unsupported and outside Homebrew's security boundary. +Running any `brew` command with `sudo` trusts the Homebrew prefix owner and anyone else who can modify Homebrew's code to run arbitrary code as root. They can change `bin/brew` or the code it loads before a later invocation with `sudo`, including before command-specific checks or switching users. This trust also applies to commands that only display information or subsequently run as another account. For example, if the owner adds `rm -rf /path/to/data` to `bin/brew`, the next `sudo brew ...` will run that deletion as root. Changing ownership of selected paths does not remove this trust in accounts that can still modify Homebrew itself. + Homebrew necessarily executes third-party software chosen by the user. Formulae and casks may run upstream build systems, installers and post-install commands. External taps, mirrors, wrappers and repositories configured by the user are trusted at the level the user granted them. Installing from a tap is a trust decision for that tap. A tap can provide formulae, casks, external commands and tap metadata that install or run software, so using a tap means trusting its authors for those files. -Homebrew's boundary for upstream software is the `brew`-managed install, post-install and test execution path. Escapes from Homebrew's intended sandboxing or credential filtering in those phases may be security issues. Once a user runs installed upstream software directly, sandboxing and trusting that software is the user's responsibility. +Homebrew's boundary for upstream software is the `brew`-managed install, post-install and test execution path. Escapes from Homebrew's intended sandboxing or credential filtering in those phases may be security issues. Once a user chooses to run installed upstream software, directly or through a Homebrew command, sandboxing and trusting that software is the user's responsibility. ## In Scope @@ -90,7 +92,7 @@ Security vulnerabilities in software used by but not written by Homebrew are not ### Local Account or Administrator Control -A report requiring the attacker to control the same user account, administrator account, shell startup files, `PATH`, Homebrew prefix, cache, lock files or repository checkout is not a security issue. Homebrew cannot defend against code already running with the user's privileges. +A report requiring the attacker to control the same user account, administrator account, shell startup files, `PATH`, Homebrew prefix, cache, lock files or repository checkout is not a security issue. Homebrew cannot defend against code already running with the user's privileges. Running `brew` with `sudo` explicitly extends trust in the prefix owner and anyone else who can modify Homebrew's code to root privileges. ### Unsupported Configurations @@ -100,6 +102,8 @@ Bugs that only affect Tier 2, Tier 3 or unsupported configurations described in Bugs in software after Homebrew has installed it are not Homebrew security issues. Homebrew may sandbox or filter credentials while it runs install, post-install and test steps, but users are responsible for deciding whether, how and where to run installed software afterwards. +This includes launching software through Homebrew commands or registering it to run automatically. Running package-provided commands with the privileges the user requested, including root, is not a Homebrew security vulnerability by itself. Homebrew's build sandbox does not confine software run this way. + ### Local Denial of Service and Resource Exhaustion Crashes, exceptions, hangs, high CPU usage, disk consumption or memory growth are not security issues when they only stop the current command or require attacker-controlled command-line input, configuration, package metadata from untrusted sources or local files. Report them publicly as quality bugs.