Repository navigation
Releases
1.56.1
1.56.0
Added
-
honeytoken plantcan now plant kubeconfig honeytokens (--type kubeconfig,--method kubeconfig). The decoy context, cluster and user are merged into~/.kube/confignext to the user's real entries and removed again when the honeytoken is revoked. On a fresh kubeconfig the decoy becomes thecurrent-context; an existing active context is never changed, so the owner's ownkubectlcannot trip the decoy. -
ggshield machine setupnow registers the machine on the GitGuardian Endpoints page, so it is listed before its first scan. -
ggshield secret scan pre-commit,pre-pushandai-hooknow send the machine hostname, from which GitGuardian registers the users of these hooks. Other scan commands and the native AI hook do not send it, and no scan sends it on CI. -
Set
endpoint_registration: falsein.gitguardian.yaml, orGGSHIELD_NO_ENDPOINT_REGISTRATIONto a non-empty value, to turn both off.
Changed
-
Commands that need git now say so in their
--help, and when git is not installed they report which command it needs, and how to install it. -
machine setupandmachine doctorno longer crash when git is not installed: the git hooks step is skipped with a warning, anddoctorreports it as skipped without failing. -
secret scan path --use-gitignorereports that it requires git instead of failing with an unexpected error. -
The Linux and macOS release binaries now bundle Python 3.14.8 instead of 3.14.6, fixing
CVE-2026-19445andCVE-2026-19553in the bundled interpreter.
Fixed
-
honeytoken plantno longer echoes a line of~/.aws/credentialswhen the file cannot be parsed. The parser quotes the offending line, so a corrupt credentials file printed the user's real secret key on stderr, and into the fleet logs when plant runs under an MDM. Only the error type and the line number are reported now. -
ggshield secret scan pre-receivenow uses the system trust store in its child process, which no longer inherited it on Python 3.14 (#1508). -
ggshield secret scan ai-hooknow verifies the instance's TLS certificate against the OS trust store. -
ggshield secret scan ai-hooknow honorsREQUESTS_CA_BUNDLEandCURL_CA_BUNDLE, from the environment or.env, like the rest of ggshield. A path that does not exist makes the scan fail, as it already does for other commands. -
On Windows,
ggshieldstarts on machines without the Microsoft Visual C++ Redistributable. Since 1.54.0 it exited silently there, because the launcher neededVCRUNTIME140.dll.install.ps1also reports an error when the installedggshieldcannot run. -
Secret scans no longer fail with a server error when a batch of files goes over the API payload limit once JSON-encoded (escaped quotes, newlines, non-ASCII or control characters). Batches are now sized by their encoded length, and a file too large to fit in any request is skipped with a message.
-
The AI hook no longer lets a tool call through unscanned when its content goes over the API payload limit once JSON-encoded: batches are sized the same way.
-
The AI hook no longer has Claude Code echo the blocked prompt back: a prompt blocked because it holds a secret used to be shown again after "Original prompt:" in the block message, secret included. The hook now sets
suppressOriginalPrompt. -
When the GitGuardian API fails with an HTML error page, for example on a scan that runs past the server time limit, ggshield now prints a one-line error with the HTTP status instead of the page. The page is still logged with
--debug.
1.56.0rc1
Added
-
honeytoken plantcan now plant kubeconfig honeytokens (--type kubeconfig,
--method kubeconfig). The decoy context, cluster and user are merged into
~/.kube/confignext to the user's real entries and removed again when the
honeytoken is revoked. On a fresh kubeconfig the decoy becomes the
current-context; an existing active context is never changed, so the owner's own
kubectlcannot trip the decoy. -
ggshield machine setupnow registers the machine on the GitGuardian Endpoints page, so it is listed before its first scan. -
ggshield secret scan pre-commit,pre-pushandai-hooknow send the machine hostname, from which GitGuardian registers the users of these hooks. Other scan commands and the native AI hook do not send it, and no scan sends it on CI. -
Set
endpoint_registration: falsein.gitguardian.yaml, orGGSHIELD_NO_ENDPOINT_REGISTRATIONto a non-empty value, to turn both off.
Changed
-
Commands that need git now say so in their
--help, and when git is not installed they report which command it needs, and how to install it. -
machine setupandmachine doctorno longer crash when git is not installed: the git hooks step is skipped with a warning, anddoctorreports it as skipped without failing. -
secret scan path --use-gitignorereports that it requires git instead of failing with an unexpected error. -
The Linux and macOS release binaries now bundle Python 3.14.8 instead of 3.14.6, fixing
CVE-2026-19445andCVE-2026-19553in the bundled interpreter.
Fixed
-
honeytoken plantno longer echoes a line of~/.aws/credentialswhen the file
cannot be parsed. The parser quotes the offending line, so a corrupt credentials
file printed the user's real secret key on stderr, and into the fleet logs when
plant runs under an MDM. Only the error type and the line number are reported now. -
ggshield secret scan pre-receivenow uses the system trust store in its child process, which no longer inherited it on Python 3.14 (#1508). -
ggshield secret scan ai-hooknow verifies the instance's TLS certificate against the OS trust store. -
ggshield secret scan ai-hooknow honorsREQUESTS_CA_BUNDLEandCURL_CA_BUNDLE, from the environment or.env, like the rest of ggshield. A path that does not exist makes the scan fail, as it already does for other commands. -
On Windows,
ggshieldstarts on machines without the Microsoft Visual C++
Redistributable. Since 1.54.0 it exited silently there, because the launcher
neededVCRUNTIME140.dll.install.ps1also reports an error when the
installedggshieldcannot run. -
Secret scans no longer fail with a server error when a batch of files goes over the API payload limit once JSON-encoded (escaped quotes, newlines, non-ASCII or control characters). Batches are now sized by their encoded length, and a file too large to fit in any request is skipped with a message.
-
The AI hook no longer lets a tool call through unscanned when its content goes over the API payload limit once JSON-encoded: batches are sized the same way.
-
The AI hook no longer has Claude Code echo the blocked prompt back: a prompt blocked because it holds a secret used to be shown again after "Original prompt:" in the block message, secret included. The hook now sets
suppressOriginalPrompt. -
When the GitGuardian API fails with an HTML error page, for example on a scan that runs past the server time limit, ggshield now prints a one-line error with the HTTP status instead of the page. The page is still logged with
--debug.
1.55.0
Added
-
ggshield now supports Amazon Kiro, both the IDE and the CLI.
ggshield machine setupinstalls the secret-scanning hooks in
~/.kiro/hooks/, and MCP servers, tool calls and past usage are reported
like they are for every other assistant. -
ggshield now supports Junie CLI, JetBrains' terminal coding agent.
ggshield machine setupinstalls the secret-scanning hooks in
~/.junie/config.json, and MCP servers and past usage are reported like they
are for every other assistant. Junie inside a JetBrains IDE is not covered:
it embeds Junie over ACP, which invokes no hooks. -
The time ggshield waits for the GitGuardian API to answer, until now fixed at 60 seconds, can be changed with the
api_timeoutconfig key or theGITGUARDIAN_API_TIMEOUTenvironment variable, which takes precedence. It accepts 1 to 3600 seconds. Raise it if a large scan fails with a read timeout.
Changed
-
The README now warns, at the top, that upgrading to 1.54.0 on macOS triggers
a Keychain prompt forggshield-py, the second binary that reads the saved
token, and that clicking Always Allow settles it for good. -
The Linux rpm now owns its bundle directories, so upgrades no longer leave
empty*.dist-infodirectories behind. -
When the AI hook finds a secret in a tool output on Claude Code or Codex, the
output is now withheld from the assistant instead of being passed to it with a
warning attached. The assistant reads the block message in place of the output,
so the secret stays out of its context and out of the session transcript on
disk, and the message no longer tells you to revoke a credential the assistant
never read. Agents with no way to replace a tool output are unchanged and keep
reporting the secret as leaked. -
On Codex, a blocked tool output was already hidden from the assistant, but the
message still said the secret had been exposed and told you to revoke it. It
now reports the output as withheld. -
Mistral Vibe behaves the same way as Codex and now reports a blocked tool
output as withheld rather than leaked. -
The Docker image now ships bytecode for the standard library, so ggshield starts
faster in containers. -
secret scan pre-receiveno longer checks for a new ggshield version unless
--check-for-updatesis passed. -
The API key check fetches server metadata and token scopes concurrently instead of
one after the other, saving a round trip on every scan. -
ggshield now requires
py-gitguardian1.35.0.
Fixed
-
Report the raw
mcp_{server}_{tool}name VSCode gives us when the server cannot
be identified from the local configuration, instead of guessing a server name that
the API cannot match. The API splits the name against the whole inventory. -
The Linux rpm package could crash on startup after an in-place upgrade,
because empty*.dist-infodirectories were left behind. They are now
removed during the upgrade. -
ggshield auth logincan now be interrupted with Ctrl+C on Windows while it
waits for the browser callback. The local callback server used to block
indefinitely, so the interrupt was only noticed once the login completed. -
The AI hook no longer rejects a hook event whose payload starts with a UTF-8
BOM, which Cursor on Windows prepends. Prompts and tool calls were reaching
the agent unscanned. -
Installing the AI hook for Codex now says that Codex will not run the hook
until it is approved in the Codex hook review, so the install no longer looks
complete while nothing is being scanned. -
ggshield machine doctorand the AI discovery no longer report an assistant as
unprotected when its hook command points at a live binary that a next upgrade
would break. Only a missing or dead command now reads as not installed, and the
command found in the settings file is reported either way. -
The AI hook no longer mistakes another assistant for Claude Code when the
transcript path merely contains "claude". A Cursor session working in a
project directory such asclaude-toolswas answered in Claude Code's output
format, which Cursor ignores, so a blocked tool call ran anyway. Every
assistant is now identified by a field only it sends, and Claude Code answers
for what remains, whatever directory it keeps its transcripts in. -
A scan request that times out waiting for the server's response now fails immediately instead of being retried for several minutes. The server has already accepted the request by the time this happens, so retrying only made it redo the same (possibly large) scan up to six times while the user waited.
-
Commit scans (
secret scan repo,commit-range,changes,ci,pre-push,
pre-receive) processed batches one at a time, ignoringGG_MAX_WORKERS.
Batches now run in parallel, which can speed up large scans considerably. -
When two ggshield AI hooks are configured for the same event, the second one no
longer exits without a verdict. Both now answer with the same decision, and the
event still costs a single scan. -
The AI hook no longer stops recognising Copilot CLI when its hook payload
carries a field ggshield has not seen before. Copilot CLI was identified by
an exact set of payload keys, so the first release adding any field would
have left its payloads claimed by no assistant at all, and an unrecognised
payload exits without a verdict and lets the tool call through unscanned. -
ggshield --no-check-for-updates <command>was ignored when the option came before
the subcommand. -
Running a command without an API key for a self-hosted instance now says how to
authenticate against that instance, instead of reportingUnknown instance. -
ggshieldno longer hangs when the working directory contains a.envFIFO
with no writer attached, as mounted by 1Password Environments. -
ggshield machine setupandggshield machine doctorno longer treat a global
or system git hook generated by an older ggshield as correctly configured. Those
hooks skip the repository's own hook in linked worktrees. Generated hooks now
carry a version stamp;doctorfails on an outdated one andsetuprewrites the
block it generated, leaving any other line in the file untouched. -
ggshield machine setupnow exits non-zero when a non-ggshield git hook is
already present, instead of reporting success on a machine with no ggshield hook. -
ggshield machine doctorandggshield machine setupnow judge only the hooks
directorycore.hooksPathactually points at. Leftover files in the directory
ggshield installs into used to sway the verdict, andsetupcould repair such a
hook and report success without ever pointing git at it.
Security
- The Docker image now installs its Python dependencies from
uv.lockinstead
of re-resolving the version ranges inpyproject.tomlat build time. The
image therefore gets the same hash-verified versions as CI, and the rolling
exclude-newer = "3 days"quarantine now applies to it too.
1.55.0rc1
Added
-
ggshield now supports Amazon Kiro, both the IDE and the CLI.
ggshield machine setupinstalls the secret-scanning hooks in
~/.kiro/hooks/, and MCP servers, tool calls and past usage are reported
like they are for every other assistant. -
ggshield now supports Junie CLI, JetBrains' terminal coding agent.
ggshield machine setupinstalls the secret-scanning hooks in
~/.junie/config.json, and MCP servers and past usage are reported like they
are for every other assistant. Junie inside a JetBrains IDE is not covered:
it embeds Junie over ACP, which invokes no hooks. -
The time ggshield waits for the GitGuardian API to answer, until now fixed at 60 seconds, can be changed with the
api_timeoutconfig key or theGITGUARDIAN_API_TIMEOUTenvironment variable, which takes precedence. It accepts 1 to 3600 seconds. Raise it if a large scan fails with a read timeout.
Changed
-
The README now warns, at the top, that upgrading to 1.54.0 on macOS triggers
a Keychain prompt forggshield-py, the second binary that reads the saved
token, and that clicking Always Allow settles it for good. -
The Linux rpm now owns its bundle directories, so upgrades no longer leave
empty*.dist-infodirectories behind. -
When the AI hook finds a secret in a tool output on Claude Code or Codex, the
output is now withheld from the assistant instead of being passed to it with a
warning attached. The assistant reads the block message in place of the output,
so the secret stays out of its context and out of the session transcript on
disk, and the message no longer tells you to revoke a credential the assistant
never read. Agents with no way to replace a tool output are unchanged and keep
reporting the secret as leaked. -
On Codex, a blocked tool output was already hidden from the assistant, but the
message still said the secret had been exposed and told you to revoke it. It
now reports the output as withheld. -
Mistral Vibe behaves the same way as Codex and now reports a blocked tool
output as withheld rather than leaked. -
The Docker image now ships bytecode for the standard library, so ggshield starts
faster in containers. -
secret scan pre-receiveno longer checks for a new ggshield version unless
--check-for-updatesis passed. -
The API key check fetches server metadata and token scopes concurrently instead of
one after the other, saving a round trip on every scan. -
ggshield now requires
py-gitguardian1.35.0.
Fixed
-
Report the raw
mcp_{server}_{tool}name VSCode gives us when the server cannot
be identified from the local configuration, instead of guessing a server name that
the API cannot match. The API splits the name against the whole inventory. -
The Linux rpm package could crash on startup after an in-place upgrade,
because empty*.dist-infodirectories were left behind. They are now
removed during the upgrade. -
ggshield auth logincan now be interrupted with Ctrl+C on Windows while it
waits for the browser callback. The local callback server used to block
indefinitely, so the interrupt was only noticed once the login completed. -
The AI hook no longer rejects a hook event whose payload starts with a UTF-8
BOM, which Cursor on Windows prepends. Prompts and tool calls were reaching
the agent unscanned. -
Installing the AI hook for Codex now says that Codex will not run the hook
until it is approved in the Codex hook review, so the install no longer looks
complete while nothing is being scanned. -
ggshield machine doctorand the AI discovery no longer report an assistant as
unprotected when its hook command points at a live binary that a next upgrade
would break. Only a missing or dead command now reads as not installed, and the
command found in the settings file is reported either way. -
The AI hook no longer mistakes another assistant for Claude Code when the
transcript path merely contains "claude". A Cursor session working in a
project directory such asclaude-toolswas answered in Claude Code's output
format, which Cursor ignores, so a blocked tool call ran anyway. Every
assistant is now identified by a field only it sends, and Claude Code answers
for what remains, whatever directory it keeps its transcripts in. -
A scan request that times out waiting for the server's response now fails immediately instead of being retried for several minutes. The server has already accepted the request by the time this happens, so retrying only made it redo the same (possibly large) scan up to six times while the user waited.
-
Commit scans (
secret scan repo,commit-range,changes,ci,pre-push,
pre-receive) processed batches one at a time, ignoringGG_MAX_WORKERS.
Batches now run in parallel, which can speed up large scans considerably. -
When two ggshield AI hooks are configured for the same event, the second one no
longer exits without a verdict. Both now answer with the same decision, and the
event still costs a single scan. -
The AI hook no longer stops recognising Copilot CLI when its hook payload
carries a field ggshield has not seen before. Copilot CLI was identified by
an exact set of payload keys, so the first release adding any field would
have left its payloads claimed by no assistant at all, and an unrecognised
payload exits without a verdict and lets the tool call through unscanned. -
ggshield --no-check-for-updates <command>was ignored when the option came before
the subcommand. -
Running a command without an API key for a self-hosted instance now says how to
authenticate against that instance, instead of reportingUnknown instance. -
ggshieldno longer hangs when the working directory contains a.envFIFO
with no writer attached, as mounted by 1Password Environments. -
ggshield machine setupandggshield machine doctorno longer treat a global
or system git hook generated by an older ggshield as correctly configured. Those
hooks skip the repository's own hook in linked worktrees. Generated hooks now
carry a version stamp;doctorfails on an outdated one andsetuprewrites the
block it generated, leaving any other line in the file untouched. -
ggshield machine setupnow exits non-zero when a non-ggshield git hook is
already present, instead of reporting success on a machine with no ggshield hook. -
ggshield machine doctorandggshield machine setupnow judge only the hooks
directorycore.hooksPathactually points at. Leftover files in the directory
ggshield installs into used to sway the verdict, andsetupcould repair such a
hook and report success without ever pointing git at it.
Security
- The Docker image now installs its Python dependencies from
uv.lockinstead
of re-resolving the version ranges inpyproject.tomlat build time. The
image therefore gets the same hash-verified versions as CI, and the rolling
exclude-newer = "3 days"quarantine now applies to it too.
1.54.0
Note
MacOS users upgrading from a previous version will be prompted with a request to grant access to the ggshield keychain secrets from ggshield-py or python. This is an intended behavior, linked to performance optimizations for AI hooks.
pip/pipxinstalls now ship the native Rustggshieldon common platforms
(Linux x86-64/aarch64 for glibc and musl, macOS universal2, Windows x86-64); the Python implementation is installed
alongside asggshield-py. Platforms without a native wheel fall back to the
pure-Python wheel, whereggshieldis the Python entry point — as does
Homebrew, which builds from the source distribution.
Added
-
Add secret-scanning hook support for Mistral Vibe 2.21 and later. Global and
project installation preserves existinghooks.tomlcontent and configures
Vibe'spre_toolandpost_toolevents. -
A native
ggshield-hookbinary implementsggshield secret scan ai-hook,
removing the interpreter startup cost the hook paid twice per tool call. It
resolves the instance's scan limits from the same places ggshield does
(GG_MAX_DOC_SIZE/GG_MAX_DOCS, the on-disk auth-check cache, then
/v1/metadata), so a document or a batch the instance would reject is split or
skipped instead of failing the scan. It caches what it learns from
/v1/metadatafor five minutes, so a hook path where only the native binary
runs does not pay that round trip on every scan. -
The native hook now honours the
GITGUARDIAN_INSTANCE,GITGUARDIAN_API_URL
andGITGUARDIAN_API_KEYsettings a.envfile provides, instead of declining
to scan when it found one. It reads the file from the same places ggshield does
(GITGUARDIAN_DOTENV_PATH, then the working directory, then the repository
root), and gives it the same precedence: a value in the.envoverrides the
same variable already exported in the environment. A self-hosted instance
configured through a.envis therefore scanned against, rather than left
unscanned. Two caveats: an unbraced$in a value is expanded as a variable
reference (soGITGUARDIAN_API_KEY=abc$defresolves toabc— quote it as
'abc$def'to keep the$), and aGITGUARDIAN_*line the parser rejects
still leaves the action unscanned, with a message naming the variable. -
AI discovery now reports, per agent, the email of the assistant subscription that agent is signed into, so a personal
subscription can be told apart from a company one. Read locally from Claude Code, Codex and Cursor. Mistral Vibe and
VSCode keep no account on disk and report nothing. -
GGSHIELD_NO_NOTIFICATIONsuppresses the AI hook's desktop notification. The
secret is still detected and the tool call still blocked; only the banner is
withheld.
Changed
-
Release binaries are now built with Python 3.14 instead of 3.10, ahead of Python 3.10's
end-of-life. This also updates the SQLite bundled in the released binaries, fixing
CVE-2025-6965andCVE-2025-7709. -
AI hooks no longer call the API twice for the same document. An unambiguously clean
scan result is cached locally for 15 minutes, keyed on the exact document sent and on
the instance and token it was sent with, so a file read costs one API round trip
instead of two (PreToolUseandPostToolUsescan the same file). -
ggshieldstarts faster: each command now imports only the modules it needs, and
plugin discovery runs only when a plugin command is actually resolved. This speeds
up every invocation, in particular the ones that run automatically on every git
commit (secret scan pre-commit) and on every AI agent tool call
(secret scan ai-hook). -
ggshield secret scan ai-hookis now much faster on MCP tool calls: the local
discovery walk is cached for an hour instead of running on every call, and the
secret scan and the MCP activity call now run concurrently instead of one after
the other. Measured against a mock API at the production p50 latency, an MCP
PreToolUseevent goes from 1183 ms to 387 ms (p50). Other events are
unaffected. As a consequence, an MCP tool call is now reported to GitGuardian
even when the secret scan blocks it. -
AI hooks now scan all the payloads of an event in a single API call instead of one
call each. A prompt mentioning three files used to cost four round trips; measured
against a mock API at the production p50 latency, such aUserPromptSubmitgoes from
1561 ms to 395 ms (p50). Single-payload events are unchanged. -
In the standalone packages (
.pkg,.deb,.rpm,.msi, Chocolatey, archives),ggshield
is now a small native dispatcher: it runsggshield secret scan ai-hookand hands every other
command to the bundled Python implementation. There is still a singleggshieldcommand, and
every command behaves as before. -
AI hooks now resolve the file paths of
Readtool calls and prompt@-mentions to an
absolute path against the event's working directory. A file mentioned in a prompt and the
same file later read by a tool now share one verdict-cache key, so it is scanned once
instead of twice. As a result, the file path reported in block messages (and sent for
scanning) is now absolute for prompt-mentioned files. -
The native hook accepts a
.gitguardian.yamlkey spelled with-or with_,
but no longer resolves a config that sets the same key in both spellings the
way the Python implementation does: the last spelling in the file wins. Using
one spelling per key, which every documented example does, is unaffected. -
When the AI hook blocks, its message now shows the
secret.ignored_matches
entry that would silence it, one line per detected secret. The message censors
every match, so the entry could not previously be written by hand from what it
showed. The value shown is the ignore sha, a digest rather than the secret,
and is the same oneggshield secret ignore --last-foundwrites. -
ggshield plugin installandggshield plugin updateno longer send an
installation report to the GitGuardian instance. The endpoint it targeted was
never implemented, so the call always failed and was silently discarded. -
The GitGuardian logo shipped with ggshield is the refreshed one: the Chocolatey
package icon uses the solid black mark, and the macOS desktop notification icon
uses the app icon on its own dark ground. -
ggshield now requires
py-gitguardian1.34.0, which carries the
AgentInfo.subscription_emailfield the AI discovery report sends.
Fixed
-
On macOS,
ggshieldno longer triggers the Xcode Command Line Tools install prompt
when the tools are not installed. -
On macOS,
ggshieldno longer pops a "Keychain Not Found" dialog when checking
whether the keyring is usable. The check now only reads from the credential
store instead of writing a probe entry, which also stops it from silently
falling back to file-based token storage on machines where the keychain can be
read but not written. -
ggshield secret scan ai-hookno longer skips the scan when a Bash command starting with
catorGet-Contentis not a plain file read (a heredoc, a redirection...).
Such commands are now always scanned as text. -
ggshield secret scan ai-hooknow scans only the lines an agent actually reads, instead of
the whole file. A file over the API's 1 MiB document limit was skipped outright — allowed
without being scanned at all — even when the agent only read a few hundred lines of it;
that slice now scans normally. Ranges are read from Claude Code (offset/limit) and
VS Code (startLine/endLine); a read without a range still scans the whole file. -
The AI discovery is now submitted to GitGuardian only when the local AI/MCP
configuration actually changed, as it was always meant to be. A type mismatch
between a freshly discovered MCP configuration and one read back from the local
cache made the comparison always report a change, so every MCP tool call
uploaded a full discovery payload. -
When a secret is found in a file mentioned in a prompt (
@path), AI hooks now name
that file in the block message instead of reporting it as a secret "in your prompt".
The previous wording pointed at the wrong content and asked the user to edit a prompt
that did not contain the secret. -
On macOS, the token stored by
ggshield auth loginis now created with a
Keychain ACL trusting every ggshield binary, so the native hook can read it
without prompting. Previously only the writing binary was trusted and each
login reset the ACL, making "Always Allow" wear off. -
On macOS,
ggshieldno longer asks for Keychain access again after every upgrade. The
code-signing identifier embedded a per-build hash, so the "Always Allow" grant recorded
for one release stopped matching the next one; it is now pinned to
com.gitguardian.ggshield. One grant made from a release built with this fix onwards
keeps working across upgrades. -
The AI hooks no longer create incidents on your dashboard when
secret.source_uuidis set
in your configuration (a setting shared with CI scanning): a hook event is not a source
scan, and every prompt and every file the agent read was creating an incident. They also
now always scan withall_secrets, so they can block on a secret that is already known to
GitGuardian instead of letting it through. -
tomliwas added to the dev dependency group, without a marker, so dev environments always have it regardless of the
Python interpreter, and ty can resolve the fallback branch it insists on analyzing -
secret scan ai-hooknow honourssecret.ignored_pathsfrom
.gitguardian.yaml, as every othersecret scancommand already did. A file
excluded there is no longer read, sent, or blocked on when an agent...
1.54.0-rc4
Pre-GA build for internal validation. Not a public release.
What changed since 1.54.0-rc3
- Mistral Vibe is recognised by the native hook.
GGSHIELD_NO_NOTIFICATIONwithholds the desktop banner, for anyone who would rather not have one.- An event whose read path is unusable no longer fails the hook open.
py-gitguardian1.34.0.- Windows wheel signing fixes, which do not affect these bundles but complete the release path.
Everything from rc2 and rc3, still in
- The AI hook no longer breaks on upgrade. The hook command points at
/usr/local/bin/ggshield, a symlink the installer repoints, instead of the versioned/opt/gitguardian/ggshield-<version>/directory an upgrade deletes.ggshield machine setupalso repairs a hook already stuck on a stale versioned path. - Commands that do not need a token (
plugin list,install,config list) no longer read the credential store at all. - A keychain read failure can no longer blank the stored token on the next save.
cryptography50.0.0. No High or Critical findings in the scan.- Refreshed GitGuardian logo on the desktop notification.
What the version numbers mean
Labelled 1.54.0-rc4, but the artifacts are built from main before the version bump:
| Where you look | What you see |
|---|---|
| This page | 1.54.0-rc4 |
| Package filenames | 1.53.0+399213e |
ggshield --version |
ggshield, version 1.53.0 |
399213e is the commit these were built from. Quote it in any bug report.
ggshield --version prints 1.53.0 on this build and on the current public release, so it cannot tell you which one you have. Ask the package manager instead:
# macOS
pkgutil --pkg-info com.gitguardian.ggshield
# Debian and Ubuntu
dpkg-query -W -f='${Package} ${Version}\n' ggshield
Both report 1.53.0+399213e when this build is installed.
Why the tag has no v prefix
Deliberate, please leave it. .github/workflows/tag.yml triggers on v* and publishes to PyPI, Docker Hub (including moving :latest), Cloudsmith and Chocolatey. A tag starting with v would ship this build to every external ggshield user. This tag fires no workflow at all.
What is in it
The native Rust AI hook. It runs on every tool call your AI agent makes, so its startup cost is paid constantly:
- macOS: about 20 ms, against about 300 ms today
- Linux: under 1 ms
ggshield is a small Rust dispatcher that answers secret scan ai-hook itself and hands every other command to ggshield-py. Everything else behaves as before.
Install
macOS (pick your CPU family), signed by GitGuardian and notarized by Apple:
sudo installer -pkg ggshield-1.53.0+399213e-arm64-apple-darwin.pkg -target /
Debian and Ubuntu (apt, not dpkg -i: the package depends on git, and dpkg will not pull it in):
sudo apt install ./ggshield_1.53.0+399213e-1_amd64.deb
The .tar.gz bundles run without installing anything.
macOS keychain
The first time something actually needs your token, macOS asks once per binary, for ggshield and for ggshield-py. Click Always Allow, not Allow: Allow grants for that one run only, so it comes back every time. Once granted it is permanent, and it survives future upgrades, because the grant is keyed on the code signature rather than the install path.
Verifying what you downloaded
shasum -a 256 -c SHA256SUMS.txt
pkgutil --check-signature ggshield-1.53.0+399213e-arm64-apple-darwin.pkg
The macOS packages report Developer ID Application: GitGuardian Inc. (N67C7J5WQ9) and Notarization: trusted by the Apple notary service.
Known issues
The bundled Python 3.14.6 carries two Low advisories (CVE-2026-18503, CVE-2026-6879), both algorithmic-complexity issues in csv and xml parsing. ggshield imports neither module. They clear when the bundled interpreter moves to 3.14.7.
No High or Critical findings.
Built by run 32373549145.
1.54.0-rc3
Pre-GA build for internal validation. Not a public release.
What changed since 1.54.0-rc1
- The AI hook no longer breaks on upgrade. The hook command now points at
/usr/local/bin/ggshield, a symlink the installer repoints on every install, instead of the versioned/opt/gitguardian/ggshield-<version>/directory that an upgrade deletes.ggshield machine setupalso repairs a hook that is already pointing at a stale versioned path, so one run fixes a machine broken by rc1. - Far fewer keychain prompts. Commands that do not need a token (
plugin list,install,config list) no longer read the credential store at all. Only something that actually authenticates will ask. - A stored token can no longer be lost. A keychain read failure used to blank the token in the config file on the next save, silently detaching it from the stored credential.
- The bundled
cryptographyis 50.0.0, clearing the High advisory reported in rc1. - Refreshed GitGuardian logo on the desktop notification.
What the version numbers mean
Labelled 1.54.0-rc3, but the artifacts are built from main before the version bump:
| Where you look | What you see |
|---|---|
| This page | 1.54.0-rc3 |
| Package filenames | 1.53.0+5a87316 |
ggshield --version |
ggshield, version 1.53.0 |
5a87316 is the commit these were built from. Quote it in any bug report.
ggshield --version prints 1.53.0 on this build and on the current public release, so it cannot tell you which one you have. Ask the package manager instead:
# macOS
pkgutil --pkg-info com.gitguardian.ggshield
# Debian and Ubuntu
dpkg-query -W -f='${Package} ${Version}\n' ggshield
Both report 1.53.0+5a87316 when this build is installed.
Why the tag has no v prefix
Deliberate, please leave it. .github/workflows/tag.yml triggers on v* and publishes to PyPI, Docker Hub (including moving :latest), Cloudsmith and Chocolatey. A tag starting with v would ship this build to every external ggshield user. This tag fires no workflow at all.
What is in it
The native Rust AI hook. It runs on every tool call your AI agent makes, so its startup cost is paid constantly:
- macOS: about 20 ms, against about 300 ms today
- Linux: under 1 ms
ggshield is a small Rust dispatcher that answers secret scan ai-hook itself and hands every other command to ggshield-py. Everything else behaves as before.
Install
macOS (pick your CPU family), signed by GitGuardian and notarized by Apple:
sudo installer -pkg ggshield-1.53.0+5a87316-arm64-apple-darwin.pkg -target /
Debian and Ubuntu (apt, not dpkg -i: the package depends on git, and dpkg will not pull it in):
sudo apt install ./ggshield_1.53.0+5a87316-1_amd64.deb
The .tar.gz bundles run without installing anything.
macOS keychain
The first time something actually needs your token, macOS asks once per binary, for ggshield and for ggshield-py. Click Always Allow, not Allow: Allow grants for that one run only, so it comes back every time. Once granted it is permanent, and it survives future upgrades, because the grant is keyed on the code signature rather than the install path.
Verifying what you downloaded
shasum -a 256 -c SHA256SUMS.txt
pkgutil --check-signature ggshield-1.53.0+5a87316-arm64-apple-darwin.pkg
The macOS packages report Developer ID Application: GitGuardian Inc. (N67C7J5WQ9) and Notarization: trusted by the Apple notary service.
Known issues
The bundled Python 3.14.6 carries two Low advisories (CVE-2026-18503, CVE-2026-6879), both algorithmic-complexity issues in csv and xml parsing. ggshield imports neither module. They clear when the bundled interpreter moves to 3.14.7.
No High or Critical findings.
Built by run 32273721253.
1.54.0-rc1
Pre-GA build for internal validation. Not a public release.
What the version numbers mean
This release is labelled 1.54.0-rc1, but the artifacts are built from main before the version bump, so:
| Where you look | What you see |
|---|---|
| This page | 1.54.0-rc1 |
| Package filenames | 1.53.0+620698b |
ggshield --version |
ggshield, version 1.53.0 |
620698b is the commit these were built from. If you report a problem, quote that.
Why the tag has no v prefix
Deliberate, and please leave it that way. .github/workflows/tag.yml triggers on v* and publishes to PyPI, Docker Hub (including moving :latest), Cloudsmith and Chocolatey. A tag starting with v would ship this build to every external ggshield user. This tag fires no workflow at all.
What is in it
The native Rust AI hook. The hook runs on every tool call your AI agent makes, so its startup cost is paid constantly:
- macOS: about 24 ms, against about 350 ms today
- Linux: under 1 ms
ggshield is now a small Rust dispatcher that answers secret scan ai-hook itself and hands every other command to ggshield-py. Everything else behaves as before.
Install
macOS (pick your CPU family), signed by GitGuardian and notarized by Apple:
sudo installer -pkg ggshield-1.53.0+620698b-arm64-apple-darwin.pkg -target /
Debian and Ubuntu (apt, not dpkg -i: the package depends on git, and dpkg will not pull it in):
sudo apt install ./ggshield_1.53.0+620698b-1_amd64.deb
The .tar.gz bundles run without installing anything, if you would rather not touch your system.
macOS keychain prompt
On first use you will be asked once for your login keychain password, from ggshield-py. That is expected: the Python half of ggshield was renamed in this version, so it needs its own keychain grant. Click "Always Allow" and it will not come back. Anything asking repeatedly, or naming something other than ggshield or ggshield-py, is worth reporting.
Known issue
The bundled cryptography 48.0.1 carries a High advisory, GHSA-g6cj-pr64-35w5, fixed upstream in 50.0.0. This is a dependency pin that predates this build and is present in the current 1.53.0 release too. It is not introduced by the Rust work.
Verifying what you downloaded
shasum -a 256 -c SHA256SUMS.txt
pkgutil --check-signature ggshield-1.53.0+620698b-arm64-apple-darwin.pkg
The macOS packages report Developer ID Application: GitGuardian Inc. (N67C7J5WQ9) and Notarization: trusted by the Apple notary service.
Built by run 32231684138.
1.53.0
Added
-
ggshield ai discover --activitynow also collects raw AI-agent activity
(transcript lines / database rows) from Claude Code, Codex, Cursor, Copilot
CLI and VSCode and ships it to GitGuardian, which scans the content and strips
secrets server-side. This is an experimental feature. -
ggshield now officially supports Python 3.13: it is covered by the CI test matrix, advertised through the trove classifiers, and
pip install ggshieldworks on a Python 3.13 interpreter. -
AI discovery now sends whether hooks are installed globally, per agent.
-
ggshield machine setupsets up all of this machine's ggshield protections in one idempotent command: the AI hook for every detected AI coding assistant, the global git pre-commit/pre-push hooks, and a honeytoken. Each protection is on by default; drop one with--no-ai-hooks/--no-git-hooks/--no-honeytokens. Narrow which assistants get the AI hook with--agent/--exclude-agent. Replaces runningggshield install -t <assistant>once per agent. -
ggshield machine setupnow installs git hooks machine-wide (for every user) when run as root or with the new--systemflag, instead of only for the invoking user. This makes MDM/fleet deployments work as expected: a single root-runmachine setupsets git's systemcore.hooksPath, so every user on the machine is covered. Without root (and without--system) it keeps installing per-user as before. -
ggshield machine doctorchecks that this machine's ggshield protections are correctly set up: the AI hooks and git hooks are installed, the GitGuardian token is reachable and carries the scopes the configured protections need (scan, plushoneytokens:writeand — when themachine_scanplugin is installed —endpoints:send, both of which require a Business or Enterprise plan), and that the plugin's native scanner loads. It is read-only, prints a specific fix for each failed check, and exits non-zero if any check fails, so it can gate an MDM rollout. -
Scan requests now include
Machine-IdandMachine-Usernameheaders -
AI scan requests now include the agent name in the headers sent to the scan route.
-
ggshield machine doctornow detects when a higher-precedence gitcore.hooksPath(a repo-local or user-global one, e.g. Husky or lefthook) shadows the ggshield git hook. Git uses only the most-specificcore.hooksPath, so such an override silently bypasses ggshield's system/global hook — doctor reports it as a failed "Git hook precedence" check (per repo) instead of giving a false sense of coverage. -
ggshield machine setupwarns at install time when acore.hooksPathoverride takes precedence in the current context, so its git hook would be shadowed there. -
New
--filename-onlyflag forggshield secret scan: when set, only the file name (not its full path) is sent to GitGuardian, so incidents record just the filename (e.g.config.pyinstead ofsrc/app/config.py). Recursive scanning still works. -
Plugins can now be installed machine-wide. Running
ggshield plugin installorggshield plugin enableas root writes to a shared system location so every user on the machine can load the plugin; non-root installs stay per-user as before. A user can still disable an admin-enabled plugin for themselves. -
install.shnow offers to add the install dir to yourPATHitself when it's missing, instead of only printing instructions: it detects bash, zsh, fish, and (independent of$SHELL) Nushell and PowerShell on Linux when installed, and updates the matching shell profile. Prompted with a default of yes, silent under-y, skippable with the new--no-modify-path.uninstall.shreverses the edit.
Changed
-
Documented the install script at the top of the README's Installation section, with the
curl | bash(Linux/macOS) andirm | iex/curl(Windows) one-liners and a pointer toscripts/install/README.mdfor the full options and uninstall. -
The documented
curl | bashinstall/uninstall one-liners no longer pass the redundant--proto '=https'and--tlsv1.2flags: the URLs are alreadyhttps://and GitHub serves only TLS 1.2+, so they added nothing (-sSfL, including-f, is kept). -
The README marks the install script as the recommended install method and shows an example that authenticates against the EU workspace or a self-hosted instance.
-
ggshield auth loginnow requests theai-discover:sendscope by default, enabling upload of AI discovery data to GitGuardian without requiring--scopes ai-discover:sendexplicitly. -
Documented the Windows MSI installer in the README's Windows installation section, with the release-page download and
msiexecinstall command. -
The Linux/macOS install script now prints shell-specific guidance, as a visible warning, when
ggshieldwon't be callable yet: either~/.local/binis not on yourPATH, or an olderggshieldinstall shadows the new one. In both cases it gives the exact line to add for your shell (zsh/bash/fish/other) plus a reminder to restart your terminal, shown at the end of the run instead of a generic note buried mid-install. -
The install-scripts README documents
-y/--yesandbash -s -- --purge -yfor unattended uninstall. -
Relaxed the upper version bounds of several dependencies (
click,oauthlib,python-dotenv,pyyaml,requestsandmarshmallow-dataclass) from the next minor release to the next major. This lets ggshield be installed alongside projects that require newer versions of these packages (for exampleclick8.2+) and lets users pull in dependency security fixes without waiting for a new ggshield release. -
ggshield machine doctornow also verifies the token carries the
ai-discover:sendscope, so a passing doctor run guarantees AI agent
discovery telemetry can be uploaded. -
AI hook block messages now clearly state whether the secret already reached the agent, and add a "How to remediate" section with numbered steps and a
ggshield secret ignore --last-foundreminder for false positives. Known secrets link back to their GitGuardian incident. The block reason is now chosen from the event (UserPromptSubmit / PreToolUse / PostToolUse) before the tool, so aPostToolUseRead or MCP call is correctly reported as an actual leak instead of the "not executed yet" wording used before the tool call. -
Updated
pygitguardianto 1.33.1, now pulled from PyPI instead of a pinned git commit.
Deprecated
ggshield install -t <assistant>(installing the AI hook per assistant) is deprecated in favor ofggshield machine setup, which configures all detected assistants at once. The command still works but now prints a deprecation notice.ggshield installremains the way to install git hooks (-t pre-commit/-t pre-push).
Fixed
-
Windows uninstaller: the
Remove the standalone ggshield at …?confirmation now shows the install path. PowerShell treated the?in$ZipDir?as part of the variable name, so the path (and the?) were dropped from the prompt. -
ggshield auth loginnow reports a clear error when the token-exchange step receives a non-JSON HTTP response, instead of crashing with an opaqueAssertionErrorand a raw traceback. -
A self-hosted instance deployed under a
gitguardian.com/gitguardian.techdomain is no longer misdetected as SaaS. Its API URL now correctly gets the/exposedprefix, soggshield auth loginand other API calls reach the backend instead of the static frontend. -
README install and usage commands no longer include the leading
$shell prompt, so copying them (via the GitHub or IDE copy button) yields a runnable command instead of$ brew install ggshield. -
A GitGuardian-hosted (SaaS) instance whose dashboard/API hostname carries a
dashboard-/api-prefix is now correctly detected as SaaS. Its API URL is built by swapping the host instead of getting a spurious/exposedprefix, so API calls reach the backend instead of the static frontend. -
ggshield api-status,ggshield auth loginand any command that reads token scopes now report a clear error when the GitGuardian API answers with a non-JSON body, instead of crashing with a rawJSONDecodeErrortraceback. This typically happens when the instance URL is wrong and the dashboard serves its HTML on a2xx. -
The install scripts no longer fail when an outdated or unauthenticated
ghcannot verify the build provenance attestation (the cause of installs aborting withunsupported tlog public key type: PKIX_ED25519on agholder than 2.56.0). Build-provenance verification is now opt-in on bothinstall.shandinstall.ps1: the install relies on the mandatory sha256 checksum by default and prints the manualgh attestation verifycommand. SetGGSHIELD_REQUIRE_ATTESTATION=1(ortrue/yes/on) to require it — that runs the check and fails the install ifghis missing, older than 2.56.0, unauthenticated, or the provenance does not verify. -
uninstall.shno longer crashes and removes nothing when run with no controlling terminal (CI, cron, MDM/launchd/systemd rollback, containers). The confirmation prompt now detects an unusable/dev/ttyand exits asking for-y, instead of aborting on an unbound variable underset -u. -
The standalone installer no longer silently ignores
--pluginwhen combined with--install-only; it warns that the plugin was not installed (because--install-onlyskips authentication) and prints the command to run later. -
An enabled plugin that fails to load (for example when a native dependency is missing) is no longer silently ignored.
ggshieldnow prints a warning naming the plugin and the reason its commands are unavailable, andggshield plugin listreports the failure status. -
ggshield config migratenow reports a clear, actionable error when no configuration file is found (for example when run from the wrong directory), instead of crashing with a raw traceback. The depr...