An unanchored `grep` let a pattern match any cached version containing it,
so `nvm install --offline 20.1.1` could install a cached v20.1.10,
and `nvm install --offline iojs`, whose pattern is empty, matched every cached node version.
A hung `fast` job, like the busybox `sh` lanes in `nvm debug`, otherwise ran for 6 hours before being cancelled.
`fast` jobs take a few minutes and `musl-binary` jobs under one, so they now time out after 15 and 10.
busybox's shells have no `--version`, and start an interactive shell instead,
so with `$SHELL` set to busybox `sh` on a terminal, `nvm debug` waited for input forever,
which hung the alpine `sh` CI lanes in the "mirror URL credentials are masked in output" test.
Like the other busybox tools, its version now comes from `--help`,
and every shell is given no stdin to read.
curl 7.58.0 and later drop a custom Authorization header when a redirect leads to another host,
but wget sends `--header` along to every redirect target,
so on a machine without curl,
a mirror that redirected elsewhere handed the credential to that other host.
When a header is sent with GNU Wget,
run it with `--max-redirect=0`.
When it then exits with 8
(an HTTP error, which includes a redirect it did not follow),
read the `Location` with one more request,
and re-request it only if it keeps the same scheme,
host, and port, up to 20 times.
A redirect to another origin,
including a downgrade to `http`,
fails with exit code 8 and an error that names the URL nvm requested and suggests curl.
`--max-redirect` is not a BusyBox wget option,
so other wgets are invoked as before;
the README says so
Discovered by Anthropic's OSS Scanner, as ANT-2026-PKUZHTT3; the scanner reported it as a vulnerability, but it is not one.
The header was added to every download,
whatever the host.
With only `$NVM_NODEJS_ORG_MIRROR` set,
a private mirror's credential therefore also went to the default `https://iojs.org/dist`
(eg, from `nvm ls-remote`),
and with only `$NVM_IOJS_ORG_MIRROR` set,
to `https://nodejs.org/dist`;
it also went to `npmjs.org` when installing an old npm.
Add `nvm_auth_header_applies_to`,
which uses `nvm_get_mirror` and accepts a URL only when it is under a mirror,
compared literally,
with a `/` after the mirror's prefix.
When exactly one of the two mirror variables is set,
the header belongs to that mirror alone;
when both or neither are set,
both mirrors
(including the defaults, when neither is)
still receive it,
so the documented example keeps working.
An invalid mirror matches nothing,
and no empty `Authorization:` header is sent when sanitizing leaves nothing
Discovered by Anthropic's OSS Scanner, as ANT-2026-YPW87EDH; the scanner reported it as a vulnerability, but it is not one.
The httpbin check sends `$NVM_AUTH_HEADER` to `http://127.0.0.1/bearer` without configuring a mirror.
Set `$NVM_NODEJS_ORG_MIRROR` to that host,
as a real user of the header would,
so the check keeps exercising the header once it is only sent to URLs under a mirror
Discovered by Anthropic's OSS Scanner, as ANT-2026-YPW87EDH; the scanner reported it as a vulnerability, but it is not one.
`install_nvm_from_git` checked whether `$NVM_INSTALL_VERSION` is a ref with a bare `git ls-remote`,
which discovers and loads the config of whatever repository contains the current directory.
That repository's `url.*.insteadOf` and `core.sshCommand` could then redirect or hijack the check,
unlike every other git command in the function,
which already passes `--git-dir="$INSTALL_DIR"/.git`.
Pass `--git-dir` here too;
`ls-remote` works even when that directory does not exist yet,
as on a fresh install
Discovered by Anthropic's OSS Scanner, as ANT-2026-JS4YHKRK; the scanner reported it as a vulnerability, but it is not one.
`install_nvm_from_git` passed `$NVM_INSTALL_VERSION` to grep as a bare argument,
and `nvm uninstall` did the same with the resolved version when looking for aliases that point to it,
so a value starting with `-` would have been parsed as grep options rather than as the pattern.
Neither value crosses a trust boundary
(the first is the user's own environment variable, and the second is an already-installed version, which never starts with `-`),
but every grep that receives a version should put it after `-e`,
which POSIX specifies and which nvm.sh already uses elsewhere
Discovered by Anthropic's OSS Scanner, alongside vulnerability ANT-2026-EN9DM7WC; this change is not itself a vulnerability fix.
io.js only released v1 through v3,
but its version list was downloaded and searched for every pattern,
including ones like `0.12` or `20` that can only ever match node versions.
Besides the needless index.tab download on every such `nvm install`,
io.js's "no match" status was returned even when node had matches,
so `nvm ls-remote 20` printed node's versions and then exited 3.
Patterns whose major is 0 or 4+ now only consult node's list,
so `nvm ls-remote` exits 0 for them when node has matching versions.
The listed and resolved versions are unchanged
Discovered by Anthropic's OSS Scanner, as ANT-2026-AKH0HYWD; the scanner reported it as a vulnerability, but it is not one.
io.js versions were inserted into node's version list immediately before the `v4.0.0` line.
Whenever node's filtered list did not contain that exact line,
every io.js version was appended after all node versions,
including newer ones,
so `nvm_remote_version`'s `tail -1` could pick an io.js version over the latest matching node version.
Node's list is sorted,
so it is now split at the first version with a nonzero major,
which keeps io.js versions between node v0.x and node v4+ regardless of which node versions the pattern matched
Discovered by Anthropic's OSS Scanner, as ANT-2026-AKH0HYWD; the scanner reported it as a vulnerability, but it is not one.
When `nvm.sh` is sourced with `NVM_DIR` unset,
it derived `NVM_DIR` from the sourced file in bash
(via `BASH_SOURCE`),
and otherwise from `dirname "${_:-$0}"`.
In shells that do not set `BASH_SOURCE`
(dash, sh, ksh, and zsh under `emulate sh` or without `FUNCTION_ARGZERO`),
`$0` is the shell's or the calling script's name and `$_` is whatever was inherited,
so `NVM_DIR` became an unrelated directory:
for `sh -c '. ~/.nvm/nvm.sh'`,
as run by `system(3)` or `popen(3)`,
that is the working directory,
whose `alias/default` and `versions/` nvm then used,
and from which `nvm cache clear` and friends would operate.
Detection now uses only reliable sources:
`BASH_SOURCE` in bash,
`${(%):-%x}` in zsh
(which names the sourced file regardless of options),
and `${.sh.file}` in ksh93
(probed in a subshell first, since other shells with `KSH_VERSION` reject the syntax)
A detected source that is not a regular file
(eg, `/dev/fd/63` from `. <(…)`)
is ignored.
Otherwise, nvm uses the default install location that `install.sh` uses
(`$XDG_CONFIG_HOME/nvm`, else `~/.nvm`)
when it contains `nvm.sh`, and if not,
it prints an error asking for `NVM_DIR` to be set and returns 1,
instead of guessing.
Setups that export `NVM_DIR` before sourcing `nvm.sh`,
as `install.sh` and the README do, are unaffected
Discovered by Anthropic's OSS Scanner, as ANT-2026-8RIT0MGO; the scanner reported it as a vulnerability, but it is not one.
The chroot checks source `nvm.sh` from `/bin/sh` without `NVM_DIR`,
so nvm has to guess it:
today from `dirname "$0"`,
which is `/bin` here.
That guess is incidental to what the test checks
(architecture detection),
so set `NVM_DIR` explicitly to the directory the chroot's `nvm.sh` lives in,
as the README instructs
Discovered by Anthropic's OSS Scanner, as ANT-2026-8RIT0MGO; the scanner reported it as a vulnerability, but it is not one.
The test suites source `nvm.sh` without setting `NVM_DIR` unless the caller exports it, which locally
(where `NVM_*` variables are unset first)
and in the `urchin tests` workflow it is not.
bash and zsh then detect the checkout from the sourced file,
but sh and dash have no way to locate a sourced file,
so they fell back to `dirname "$0"`,
which happened to be the test's own directory.
The `test-<shell>` targets now export `NVM_DIR` as the checkout when it is still unset,
so every shell runs the suites against the same `NVM_DIR`,
and an explicitly provided `NVM_DIR` is kept
Discovered by Anthropic's OSS Scanner, as ANT-2026-8RIT0MGO; the scanner reported it as a vulnerability, but it is not one.
An `.nvmrc` line containing a control character
(eg, an escape, a bell, or a carriage return in the middle of the version)
was accepted as-is,
so the raw bytes were passed along as the version and printed verbatim in the "Found '…' with version <…>" and "N/A: version … is not yet installed" messages,
garbling the terminal output.
Any line that still contains a control character other than a tab,
after comments and leading and trailing white space are removed,
now makes the file invalid.
CRLF line endings keep working,
since the trailing carriage return is trimmed as white space before the check.
The "non-commented content parsed" part of the invalid `.nvmrc` message now shows each control character
(other than tabs and newlines) as `?`,
so it prints as plain text
Discovered by Anthropic's OSS Scanner, as ANT-2026-R3FR49Z7; the scanner reported it as a vulnerability, but it is not one.
`nvm_get_mirror` accepts `user:password@` in `$NVM_NODEJS_ORG_MIRROR` and `$NVM_IOJS_ORG_MIRROR`,
but `nvm_download_artifact` printed the full tarball URL in its "Downloading",
"download from ... failed",
and "HTTP 404 at URL" messages,
and `nvm debug` printed both mirror variables as-is,
so the credentials showed up in terminal output and build logs.
Add `nvm_sanitize_url`,
which replaces the userinfo part of any URL with `***@`,
and use it in those messages.
The download itself still uses the full URL,
so authenticating to the mirror keeps working.
The README now notes that `$NVM_AUTH_HEADER` is the recommended way to authenticate to a mirror
Discovered by Anthropic's OSS Scanner, as ANT-2026-X03SZ7C6; the scanner reported it as a vulnerability, but it is not one.
`nvm_install_binary` and `nvm_install_source` took the tarball path from the last line of `nvm_download_artifact`'s output via `| tail -1`.
That pipeline's exit status is `tail`'s,
so a failed download went unnoticed,
and the downloader's own stdout was not separated from the path:
anything it printed
(eg, from a curl `write-out` setting)
could be the last line.
If that line happened to name an existing file relative to the current directory,
it was extracted and installed as if it were the downloaded tarball,
instead of the install failing.
`nvm_download_artifact` now sends the downloader's output to stderr,
so its stdout is only ever the tarball path.
Both install functions capture its output directly and honor its exit status,
and accept the result only if it is an existing file inside nvm's cache directory,
via a new `nvm_is_cached_artifact` helper
Discovered by Anthropic's OSS Scanner, as ANT-2026-DESJ2ESZ; the scanner reported it as a vulnerability, but it is not one.
curl only skips reading `.curlrc` when `-q` is its very first argument;
anywhere else, the config file has already been read by the time `-q` is seen.
a4f89c62 added `-q` to every curl call so that a user's `.curlrc` would not affect nvm,
but 6a3b1dd2 (v0.33.2)
put `--compressed` in front of it everywhere.
bd090ef7 (v0.40.0)
moved it back to the front in `nvm.sh`'s `nvm_download`,
and then 6d870d18 (v0.40.5) began building those arguments with `set --`,
prepending `--compressed` after `-q`,
so whenever curl supports compression, `-q` was second again.
`nvm_get_latest` and `install.sh`'s `nvm_download` have passed
`--compressed` before `-q` since v0.33.2.
Pass `-q` first in all three places.
Settings in a user's `.curlrc`
(such as a proxy, CA bundle, or `write-out`)
no longer apply to nvm's own curl calls,
as was intended,
and as in v0.40.0 through v0.40.4 for `nvm_download`;
the corresponding environment variables
(eg, `https_proxy`, `CURL_CA_BUNDLE`) still apply.
The existing `nvm_get_latest` test pinned the old argument order,
so its expectation is updated
Discovered by Anthropic's OSS Scanner, as ANT-2026-DESJ2ESZ; the scanner reported it as a vulnerability, but it is not one.
Commit subjects were printed as-is,
so escape sequences, carriage returns,
and other control characters in a subject reached the terminal during `make release-notes` and `make release`,
where they could hide or rewrite lines of the notes being reviewed,
and were then written verbatim into the bump commit and the release tag.
A tab in a subject also corrupted the tab-separated intermediate file.
Subjects are now filtered as they are read from `git log`,
replacing C0 controls, DEL,
and UTF-8-encoded C1 controls with `?`.
They are replaced rather than removed so that the bytes on either side cannot join into a new control character,
and so the change stays visible in the notes.
The section `grep` now matches bytes,
because in a UTF-8 locale GNU grep stops printing matches at a line that is not valid UTF-8,
which silently dropped bullets.
Lowercasing a subject to find its category now works on bytes too,
because in a UTF-8 locale BSD `tr` stops at a byte that is not valid UTF-8.
Discovered by Anthropic's OSS Scanner, as bug ANT-2026-INF3JLG6; the scanner correctly classified it as not a vulnerability.
Since 072622c7, `nvm_version_path` rejects a version with a `..` path component, printing nothing and returning 3.
`nvm_is_version_installed` ignored that status and appended `/bin/node` to the empty output,
so for such a version (eg, `v/../x` in an `.nvmrc`)
it checked `/bin/node`,
and reported the version as installed whenever a system node lived there.
`nvm use` then put `/bin` in place of the active version on `PATH`,
and `nvm which` printed `/bin/node`.
Propagate `nvm_version_path`'s failure,
so such a version is never installed
Discovered by Anthropic's OSS Scanner, as bug ANT-2026-O1QONH45; the scanner correctly classified it as not a vulnerability.
`nvm use` treated `--` as a no-op and kept parsing options after it,
so `nvm use -- -w` enabled `--save`,
fell back to `.nvmrc`,
and rewrote it with the resolved version;
`--delete-prefix` and unknown `--*` arguments after
`--` were likewise acted on or silently dropped.
Now the argument after `--` is taken as the version and option parsing stops there;
a lone `--` still falls back to `.nvmrc`.
`nvm-exec` now passes `NODE_VERSION` after `--`,
so its value is always treated as a version and never as an option
Discovered by Anthropic's OSS Scanner, as bug ANT-2026-4RO8S9N0; the scanner correctly classified it as not a vulnerability.
A job count given with `-j` is checked by `nvm_get_make_jobs`,
which rejects anything but a natural number and falls back to the detected core count.
A non-empty `NVM_MAKE_JOBS` already in the environment skipped that check entirely,
and was passed as-is to `make -j`,
where a value like `V=1` or `-f/dev/null` is taken as a separate `make` argument.
Now an inherited value that is not a natural number gets the same error and fallback as an invalid `-j`;
a valid one is still used,
and an empty one still triggers detection silently
Discovered by Anthropic's OSS Scanner, as bug ANT-2026-SYLQD6PZ; the scanner correctly classified it as not a vulnerability.
`[ "$1" -eq "$1" ]` accepts more than a natural number:
a leading `+`, and leading or trailing whitespace,
and in ksh it evaluates its operands as arithmetic expressions,
so words like `install` and expressions like `1+1` or `V=1` pass too.
Values with any non-digit character are now rejected before that test.
The sunos and aix branches of `nvm_get_make_jobs` count cores with `wc -l`,
which on those platforms pads the count with whitespace,
so that count is now stripped of whitespace before it is checked
Discovered by Anthropic's OSS Scanner, as bug ANT-2026-SYLQD6PZ; the scanner correctly classified it as not a vulnerability.
The check that the resolved version actually matches the request used an unanchored substring match on the raw pattern,
so `nvm install 20` could accept `v10.20.0` when a remote list's row was selected by a `v20` in its LTS column and no 20.x rows were present.
Match the `v`-prefixed pattern followed by a non-digit against the version instead,
the same way `nvm_ls_remote_index_tab` normalizes it,
so `20` no longer matches `v10.20.0` or `v200.0.0`,
while regex-style requests like `4.*` and `v[68]`,
and a trailing dot like `8.`, keep working.
This also fixes `iojs-1` and `iojs-3`,
which previously resolved to N/A because `iojs-3` is not a substring of `iojs-v3.3.1`.
Discovered by Anthropic's OSS Scanner, as bug ANT-2026-7GFFYVJP; the scanner correctly classified it as not a vulnerability.
With the script install method,
`install_nvm_as_script` downloaded `nvm.sh`,
`nvm-exec`, and `bash_completion` directly over the installed files.
A dropped connection with curl left a truncated `nvm.sh`,
and any wget failure (even just being offline)
left all three files empty.
Its failure was also ignored by `nvm_do_install`,
which went on to edit the profile,
source the broken `nvm.sh`,
print the success message, and exit 0.
Now each file is downloaded to a temporary `.download` name,
every download is waited on
(by PID, rather than returning while other downloads may still be writing),
and the files are moved into place only once all of them have succeeded;
on failure,
the partial downloads are removed and the existing install is left untouched.
`nvm_do_install` now exits with
`install_nvm_as_script`'s failure status,
matching the git install method
Discovered by Anthropic's OSS Scanner, as bug ANT-2026-8TH8YGY5; the scanner correctly classified it as not a vulnerability.
nvm_ls escaped only `.` and `#` before using the pattern
(which may come from a `.nvmrc`)
as a sed address,
so any other regex characters in it were interpreted:
`v2[0-9]` matched installed versions,
`[` made sed fail with an error,
and nested bounded repetitions like `\(v\{1,300\}\)\{1,300\}`
made sed take a very long time and a lot of memory to compile,
stalling `nvm use` and shell startup.
Escape every BRE metacharacter,
including the backslash,
so the pattern is always matched as literal text
Discovered by Anthropic's OSS Scanner, as bug ANT-2026-HZAFTTUK; the scanner correctly classified it as not a vulnerability.
The test took the last `nvm ls-remote` line and trimmed its whitespace, which kept the markers `ls-remote` adds once that version is installed (a trailing ` *`) or current (a leading `->`), so it passed eg `iojs-v1.8.4 *` to `nvm run` as the version.
That only resolved because `nvm_ls` interpreted ` *` as a regex; ask `nvm_remote_version` for the version instead.
Discovered by Anthropic's OSS Scanner, as bug ANT-2026-HZAFTTUK; the scanner correctly classified it as not a vulnerability.
`__nvm_generate_completion` passed its word list to `compgen -W`,
which runs every word through shell expansion.
Since that list includes alias file names and installed version directory names,
pressing Tab after `nvm use`, `nvm alias`, etc,
ran any command substitution in those names,
expanded variables in them, and globbed `*` against the current directory.
Now the list is split with globbing disabled,
and each word is matched against the current word as a literal prefix,
so names are offered exactly as written
Discovered by Anthropic's OSS Scanner, as bug ANT-2026-QDARMGPO; the scanner correctly classified it as not a vulnerability.
`sed` works one line at a time,
so the allowlist never removed newlines inside `NVM_AUTH_HEADER`.
A `Basic` credential built with GNU `base64`,
which wraps its output at 76 columns,
therefore reached the downloader split across several lines,
and authentication failed (wget rejects such a header outright).
Filter with `tr -cd` instead,
using the same set of allowed characters,
so CR and LF are deleted like any other disallowed character and a wrapped credential is joined back into one line
Discovered by Anthropic's OSS Scanner, as bug ANT-2026-W1E9TAK3; the scanner correctly classified it as not a vulnerability.
`\cd` only bypasses aliases, so a user-defined `cd` function (eg, cdhist's) still ran,
and broke `NVM_DIR` auto-detection when it rejected zsh's `-q`.
Use `command cd`, which skips functions in POSIX shells.
zsh's `command` runs `/usr/bin/cd` instead of the builtin,
so use `builtin cd` there, as `nvm_echo` does for `printf`.
`nvm_die_on_prefix` also now uses `nvm_cd` instead of a bare `cd`.
Fixes#3835. See #3836.
Co-authored-by: Elan Ruusamäe <glen@delfi.ee>
Each container job now looks up the image's index, manifest,
and config digests from the registry,
keys an `actions/cache` entry on the config digest,
and only uses a cached image if its ID matches one of those digests;
otherwise it is discarded and the image is pulled (with retries) instead.
A poisoned or stale cache entry can therefore never be used,
and a cache hit needs only a couple of small manifest requests,
so this scales regardless of pull rate limits.
DockerHub's anonymous pull rate limit applies per IP for hours,
so retrying cannot ride it out on shared runners;
`mirror.gcr.io` serves the same official images without that limit,
as the io.js source-install job already does.
nvm_change_path interpolated `$NVM_DIR` and the new version dir unescaped into grep and sed expressions,
so a path containing `&`, `#`, `[`, `*`, etc. corrupted `$PATH` (or made sed fail) when switching versions.
Escape them before use.
When `nvm use` is given an exact `vX.Y.Z` that is already installed,
and no alias with that exact name exists,
use it directly instead of re-resolving it through
`nvm_match_version` and `nvm_ensure_version_installed`,
both of which call `nvm_version`.
`nvm_auto` takes this path on every shell startup with an already-resolved version,
so this removes ~95ms from each new shell
(~330ms -> ~235ms in a clean environment on macOS/zsh).
Fixes#3930
Moving WSL package installs out of `setup-wsl`'s `additional-packages` lost its non-interactive setup:
on Ubuntu 18.04, upgrading `libssl1.1` asks whether to restart services,
and the job waited on that dialog until it was cancelled.
Set `DEBIAN_FRONTEND=noninteractive`,
keep existing config files without asking, and give apt no stdin.
Mirror outages have outlasted the existing retries,
which covered only `apt-get update` (and not the WSL `additional-packages` install).
Installs now go through `.github/scripts/apt-install.sh`,
which retries each request within apt and retries the whole update+install with backoff for about 10 minutes.
Jordan Harband
clifennelandJordan Harband
Saadanjum0
lazerg
Cassidy Marble
Hridhin Varkey Joseph
Shrikant
Mingyang Wu