Aller au contenu principal

2026.9

Versioncalendar icon Release Date
2026.9.0September 25, 2026

System Requirements Update​

Ensure your infrastructure meets the latest requirements for optimal performance and security:

ComponentMinimum VersionRecommended Version
KOTS1.117.3Latest
Kubernetes1.301.35
PostgreSQL1517
Redis67
ggscout0.32.0Latest

The ggscout minimum moves to 0.32.0, the first version that classifies collected secrets with the Secrets Detection Engine. The chart ships 0.32.0; upgrade any ggscout you run outside the cluster.

Helm & Upgrade Considerations​

To ensure compatibility, please review Helm values updates from the previous version. Air gap deployment? Find all the images and tag names in the air gap install page.

Upgrading to 2026.9

Jira Data Center and Confluence Data Center: this release removes the system that renewed Personal Access Tokens on your behalf. Before upgrading, update the token of every Jira Data Center and Confluence Data Center integration, sources and notifiers alike, with a long-lived PAT. Update it in place from the integration settings, do not reinstall. An integration still relying on renewal stops working when its token expires. See the customer notice.

ClickHouse (Helm installations): ClickHouse is now rendered by the GitGuardian chart itself, under a single clickhouse.* key, instead of the vendored clickhouse-server subchart. If you enabled ClickHouse in 2026.8, move your clickhouse-server.* values to clickhouse.* before upgrading: the chart refuses to render while a clickhouse-server block is present. Object storage and backup settings are now structured under clickhouse.objectStorage and clickhouse.backup, and the clickhouse.serverConfig keys switch to camelCase. See ClickHouse storage and the Helm values changelog.

Argo CD installations: this release only creates the proxy Secrets, gim-proxy and gim-migration-proxy, when you actually set a proxy URL. Earlier releases created them on every install. If you do not use an outbound proxy, upgrading therefore leaves the old Secret behind, and Argo CD does not remove resources a chart has stopped rendering unless pruning is enabled. The sync ends with 1 resources require pruning and the application stays OutOfSync, even though its health is Healthy. Nothing is broken meanwhile, because the Secret is only ever read as an optional value. To settle the sync, prune once, either by enabling automated pruning on the application or by running argocd app sync <app> --prune.

Feature highlights​

  • GitGuardian Bridge on Self-Hosted: scan sources that live in an isolated network, such as a GitLab instance behind a firewall, without opening an inbound connection into it. You install the bridge server yourself, declare each bridge in your Helm values under GGBridge.bridges, and the bridge client inside the isolated network dials out to it. Learn more.
  • Honeytoken protection on endpoints: ggshield plants a decoy AWS credential in the file credential-harvesting malware reads first, so any use of it raises an alert naming the compromised endpoint. A coverage card in the Endpoints dashboard tracks the share of your fleet that is protected. Requires Endpoint Protection, and therefore ClickHouse, on your instance. Learn more.
  • Vaulted secrets classified by value: ggscout now runs the Secrets Detection Engine on every value it collects, so the NHI inventory can filter by detector, detector type, secret family, category, and provider. Requires ggscout 0.32.0 or later. Learn more.
  • AI Hooks for Codex and Mistral Vibe, and the AI agents and MCP inventory: AI Hooks now block secrets inside Codex and Mistral Vibe, alongside Claude Code, Cursor, Copilot CLI, and VS Code. The new inventory shows which AI agents and MCP servers run on each endpoint, whether hooks are installed, and whether an agent runs under a personal or a company subscription. The inventory is unlocked by the Endpoints entitlement of your license and requires ggshield 1.54.0 or later. Learn more.

Secrets Detection Engine​

  • v2.170: 6 new detectors (Firebase ID Token, Coze Access Token, Keycloak Access Token, Keycloak Refresh Token, Fish Audio API Key, Azure Custom Vision Training Key), 8 improved detectors (Fastly Personal Token and Scaleway Token precision, Docker Hub personal access tokens used with docker login --username, revoked-key detection for the three New Relic checkers, Bitbucket Keys checker no longer reports invalid credentials as valid, Google OAuth2 Keys banlist), 1 new analyzer (Apify Token), 1 analyzer fix (Google Cloud Keys).
  • v2.171: 7 improved detectors (n8n API Key and n8n MCP Access Token host detection, Hashicorp Vault AppRole Authentication recall, Generic High Entropy Secret no longer reports variables ending in checksum, fixed Scaleway Token, Workato API Key and Open VSX Access Token checkers), 2 new analyzers (Akamai API Credentials, Docker Credentials), 3 analyzer updates (GitLab Token, Figma Personal Access Token, OpenAI API Key).

Enhancements​

  • SSO and SCIM sign-in now works when the identity provider supplies no first and last name, which are derived from the user's email address instead. Learn more.
  • Linked Jira tickets get comments for more incident events (reassignment, regression, validity, access grants and revocations, public sharing, feedback), per-team incident permissions can be set when promoting a user to Member, and the unhealthy integration banner shows only to managers and owners. Learn more.
  • Self-Hosted:
    • ClickHouse is now covered by the Admin Area Health check page, the admin banner, and the notification emails: reachability over HTTP, metadata volume fill level against a configurable threshold, and backup freshness read from the backup sidecar. When ClickHouse is down, only reachability turns red and the other two checks report the reason instead of guessing. Its logs and collectors are included in the support bundle, and the ClickHouse and backup sidecar Prometheus metrics come with recommended thresholds and alerts. ClickHouse remains optional in 2026.9. Learn more or see the health check page.
    • The in-app analytics job can write to PostgreSQL schemas you choose, through inAppAnalytics.postgresql.applicationSchema, useSeparateAnalyticsSchema and analyticsSchema, for databases that forbid the public schema or cannot grant CREATE to the GitGuardian user. Leaving them unset keeps the current layout. Learn more.
    • The support bundle pod can be scheduled with replicated.supportBundle.nodeSelector, tolerations and affinity, for example on tainted nodes, and the Admin Area now reports why a support bundle pod cannot be scheduled instead of waiting silently.
    • Your own MCP gateway can be registered as an OAuth callback through mcpServer.oauth.extraRedirectUris.
    • GitHub app installations made outside the dashboard, for example from the app's public GitHub page, no longer join the monitored perimeter on their own. They wait on the GitHub or GitHub Enterprise Server settings page until an Owner or a Manager approves or rejects them. Installations started from the dashboard are unaffected.
    • Validity checks can leave a detector's default host out and check secrets against your custom hosts only, which matters on air-gapped networks that cannot reach the vendor's public host. Learn more.

Fixes​

  • Invalid incidents that a playbook filter should have ignored are now ignored, missing scan statuses were added to the pygitguardian client and the API documentation, Jira Cloud recurrent scans no longer fail on multi-page issue listings, and an integration no longer remains out of sync after a GitHub outage. Learn more.
  • The revoke button is displayed again for revocable secrets to members with Full Access. Learn more.
  • Uninstalling a GitLab integration no longer fails. Learn more.
  • Source locations are no longer recomputed too frequently, which caused excessive repository clones and bandwidth usage. Learn more.
  • Self-Hosted:
    • ClickHouse pods failed to start with ImagePullBackOff because the volume-permissions init container image could not be pulled. The init container now uses the wolfi/bash image the chart already ships.
    • The MCP server image is pulled through the Replicated proxy like every other image, so proxy-only installs no longer need direct access to ghcr.io.
    • The chart no longer creates proxy Secrets on installs that set no proxy URL.
    • Loki queries no longer miss log lines. After a log rotation, gunicorn workers could keep writing to several log files at once while fluent-bit tailed only one of them. The log file is now rotated once across all forked workers.
    • Triggering an ML backpopulate scan from the Admin Area gives immediate feedback again, and the same run can no longer be started twice. This was a 2026.8.0 regression.
    • Endpoint scans sent by a newer ggshield are accepted even when they carry fields the server does not know yet, instead of being rejected.