Fides 2.90.0 release notes

• 2.90.0: July 20, 2026

Fides 2.90.0

Helios

Detect and classify global data risks with comprehensive visualization for data mapping and reporting for modern Enterprise.

• Entra Monitor Is Now Generally Available: The Microsoft Entra IdP monitor is no longer behind a feature flag. Enrichment defects were also fixed — windowsazure.com leakage, an LLM-classification audit-trail wipe, and post-LLM Compass re-enrichment ignoring the LLM-supplied domain.
• AWS Monitor Refresh: The AWS monitor Action Center gains bulk actions and an updated screen experience aligned with the latest designs.
• Website Monitor PDF Compliance Reporting: The Website Monitor PDF report now includes compliance data, backed by new consent-summary and consent-issues report endpoints.
• Multi-Select Taxonomy Custom Fields: Taxonomy-backed custom fields can now be multi-select, using a [] suffix convention on field_type (mirroring the string vs string[] pattern) — no database migration required.
• Batch LLM Classification: Datastore classification can now classify a table's fields together in parallel LLM calls, speeding up classification of wide tables.
• Faster System Configure Page: The system configure page now loads in well under a second (previously several seconds). The Compass vendor list is served from the local cache instead of proxying to Compass on every request, and the page no longer blocks rendering while the vendor dictionary downloads.
• Full Classification Rationale: The LLM classification rationale is no longer truncated on staged resources — data stewards see the complete reasoning behind each classification. Includes a database migration (drops an unused index that imposed a row-size limit).

Lethe

Enterprise grade privacy request management and processing with a single orchestration layer for DSRs across vendors and systems

• California Privacy DROP Connector: Building on the 2.89.0 foundation, the California Privacy DROP connector now ships end to end — connector configuration UI (erasure policy, list types, environment), gather-query mapping and backfill tabs, a cycles UI with live erasure progress and response-file downloads, a privacy-request DROP tab with source badges, and the full admin API (connector CRUD, cycle history/detail, on-demand run-cycle/sweep, audit log). Under the hood it adds identifier normalization with Snowflake SQL parity, a match index with scheduled incremental backfill, a streaming cycle orchestrator, status derivation with idempotent privacy-request minting, a suppression ledger, and tiered response-file upload/rejection handling. Append-only enforcement on DROP tables now uses database triggers. Includes database migrations.
• DSR Traversal Preview Visualizer: The DSR traversal preview gains a full visualizer — reachability reasons, node-status markers, and manual-task detail; plain-English explanations of why an integration is unreachable or needs a manual identity; disabled, graph-build-failed, skip-processing, and empty integrations shown as marked nodes instead of being dropped; access/erasure derived from a selected policy; and on-demand per-field detail. Load times improved dramatically by caching pre-built dataset graphs and a background-prewarmed materialized graph store, with a refresh endpoint to force a recompute after datasets change and runtime telemetry for observability.
• Manual Task Enhancements: Manual tasks gain upstream connector context for reviewers, an optional link field for task-fulfillment URLs, and a stable sort order so rows no longer reorder while tasks are completed.
• Redaction Event in the Activity Timeline: The access-package redaction event now appears on the privacy request activity timeline with a dedicated icon, "Redacted" label, and status color. The timeline itself was refreshed with integration logos, friendly connection names, leading row icons, and clearer event labels.
• Single-Rule Manual-Only Policies: Manual-only request types (correction, restrict-processing, automated-decision, complaint) are now constrained to a single rule per policy, matching how the manual-only pipeline runs.
• Bug Fixes:
  • Stripe and Stripe (Organization) erasure ordering was fixed so cards are deleted after payment methods are masked/detached, preventing 404s that left payment methods unmasked.
  • Email identities are normalized to lowercase at ingestion for case-insensitive matching.
  • Request types now use consistent, human-readable display names across the admin UI.
  • Jira tickets are no longer falsely marked deleted when an invalid credential degrades to anonymous access; the credential failure surfaces on connection health instead.
  • Fixed a connection leak in the digest scheduler that could exhaust the database connection pool.
  • The DSR traversal preview now refreshes correctly after a dataset is edited via the API or SDK. Includes a database migration.

Janus

High performance consent recording and orchestration for data sharing, built for enterprise data engineering and AI pipelines.

• GPP DisclosedVendors Segment: The generated GPP string now includes the DisclosedVendors segment in the TCF EU section for TCF v2.3 compliance.
• Faster v3 Consent Reads: The v3 privacy-preferences list and create endpoints drop redundant ORM usage and duplicate validation passes for lower-latency consent reads and writes.
• Bug Fixes:
  • Fixed TCF "Reject all" rejecting legitimate interests when clicked before the full experience loaded under the reject-consent-only mechanism.
  • Stopped a fides_string NoticeConsent override from re-saving identical consent to the privacy-preferences API on every page load.
  • The Consent Form modal now hides read-only notice toggles entirely instead of showing them disabled.

Astralis

Real-time data access, usage, and retention policy enforcement across your infrastructure. Astralis embeds governance directly into data pipelines and AI workflows, preventing misuse before it happens and generating an always-on audit trail for regulatory and AI governance.

• PBAC-Gated MCP Gateway: Fides can now act as a PBAC-gated MCP gateway. Operators register upstream MCP servers, and every tool call is evaluated by the PBAC engine before it reaches data — denied calls are blocked and recorded on the Access Control page. It includes a self-hosted OAuth 2.1 authorization server (dynamic client registration, PKCE, refresh) so MCP clients log in via browser. A companion gateway mounted at /mcp exposes the Fides API itself through three meta-tools (search, describe, call) that run under the caller's own RBAC, with confirm-before-write guidance and sensitive routes (subject PII, account/auth records, delivery-channel credentials) excluded. Disabled by default (FIDESPLUS__MCP_GATEWAY__ENABLED). Includes a database migration.
• Create Assessments From Slack: You can create privacy assessments from a Slack conversation with Astralis — @mention it in any channel or DM — using a ReAct loop to gather the template and system (adding the system to the data map if needed), then posting a deep link and autofilling answers in the background.
• Dynamic Follow-Up Questions: When the last main questionnaire question is answered, an end-of-round evaluator reviews the complete answer set and asks up to three follow-up questions for uncaptured disclosures, cross-group implications, or contradictions before the questionnaire completes. The questionnaire chat also moved from a two-pass LLM pipeline to a single agentic tool-use loop for more natural, context-aware replies.
• Styled Assessment PDF Report: The privacy assessment report was redesigned as a styled, client-side PDF export matching the website monitor report.
• Resilient Assessment LLM Calls: Privacy-assessment LLM calls now retry on transient provider errors (rate limits, 5xx, timeouts) with exponential backoff.
• Bug Fixes:
  • Privacy assessments are now correctly gated behind privacy_assessment:read (fixes a 401).
  • Agent-generated assessment answers are no longer attributed to the requesting user, and system-derived answers are consistently labeled "Agent."

Integrations

• Iterable Lists Collection: The Iterable connector gains a lists collection for access and erasure.
• Bug Fixes:
  • Fixed AWS Bedrock connection tests ignoring the requested region when no explicit credentials were supplied.
  • Retagged Jira avatarUrls fields from user.biometric to user.content.

Fides core

• Multi-Org Management (Preview): A self-referential organization hierarchy on ctl_organizations plus an RBAC-scoped Organizations management page — browse the org tree, edit/create any organization, re-parent, and set the default. Organization endpoints are scoped to the caller's org grants. Feature-flagged off by default (multiOrgManagement). Includes a database migration.
• Stricter Database Health Reporting: GET /health/database now returns 503 when the database needs migration or is unreachable, so orchestrators don't route traffic to a webserver whose schema is behind its code. A database migrated by a newer build (code behind database) still returns 200, since the schema is backward compatible.
• Event-Loop Responsiveness: Token-endpoint credential hashing now runs in the threadpool so logins no longer block the event loop, and garbage-collection pauses that stalled the loop under load can be reduced via performance.freeze_gc_at_startup.
• Safer Concurrent Migrations: Database migrations are serialized behind a Postgres advisory lock so concurrent webserver pods no longer race to create the alembic_version table on a fresh deployment.
• Form Builder Resilience & Accessibility: The Privacy Center form builder handles partial/aborted LLM streams by restoring the previous form state, and gains accessibility and performance improvements.
• Ant Design Migration: Continued migration of admin-ui tables, modals, and lists from Chakra UI to Ant Design (System Resources, Consent Management vendors, Dataset Management, confirmation dialogs, the ScrollableList component, checkbox-tree filters to multi-select dropdowns, and remaining hand-rolled tables), plus an upgrade to Ant Design 6.5.0 and follow-up polish (collapsed nav icons, borderless Select focus rings, blank access-policy canvas, name-based dropdown filtering).

Security

• Dependency and Image Remediations: Bumped pydantic-settings to 2.14.2 (GHSA-4xgf-cpjx-pc3j), msgpack to 1.2.1 (GHSA-6v7p-g79w-8964, a use-after-free DoS), ujson to ~=5.13.0 (CVE-2026-54911), and weasyprint to 69.0 (CVE-2026-49452, CSS injection). Bumped 12 JS/TS dependencies to clear npm audit advisories, and removed build-only npm dependencies and the npm CLI from production Docker images.

Database schema & data changes

• This release includes multiple database migrations. Notable ones: the full California Privacy DROP connector (models, suppression ledger, append-only triggers), the multi-org organization hierarchy, manual-task context sources, dynamic follow-up questions, the MCP gateway, DSR traversal-preview cache invalidation, and dropping an unused GIN index so classification rationale is no longer size-limited.

Notes

• The Entra IdP monitor is now generally available — its feature flag has been removed.
• GET /health/database behavior changed again: it now returns 503 when the database needs migration or is unreachable (reversing part of the 2.89.0 change), while a database migrated by a newer build still returns 200. Update readiness/liveness probes accordingly.
• Health-check access logs moved to DEBUG: /health and its subpaths are now access-logged at DEBUG instead of INFO, keeping probe traffic out of production INFO logs.