Fides 2.92.0 release notes

• 2.92.0: August 17, 2026

Fides 2.92.0

Helios

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

• Microsoft Purview Discovery: Microsoft Purview can be added as an integration and monitored for datastore discovery, inventorying the databases, schemas, tables, and columns it has already catalogued without connecting to each underlying datastore. Behind the purviewIntegration feature flag. Includes a database migration.
• Batch Classification In The Monitor Form: The per-monitor batch-mode LLM classification setting is exposed in the datastore discovery monitor configuration form, so batch mode can be turned on per monitor from the UI.
• PII-Safe Dataset Testing: SaaS connector datasets can be tested without exposing personal data — results are metadata-only by default, raw data access is gated, async DSRs are blocked during testing, and access is audit logged. The results panel adds a data categories view, a syntax-highlighted JSON editor, and a tab-aware copy button.
• External References In The Nodegraph: The dataset nodegraph surfaces external field references and links them to the dataset they point at.
• Cloud Infrastructure Error History: Failed transitions on cloud infrastructure staged resources are recorded in a dedicated error table, and cloud infra resource API responses now include an errors list. Includes a database migration.
• Bug Fixes:
  • Systems can be assigned to cloud infrastructure resources that are in a promotion error state.
  • The Action Center "Categories of consent" columns and filters show every data use, with recommended categories ordered first.
  • Selecting a row in the Action Center infrastructure systems table is no longer silently cleared by an unrelated re-render.
  • The dataset editor no longer drops field or collection metadata on save, and its internal links to dynamic routes resolve correctly.
  • Protected-field decorations no longer disappear in the dataset YAML editor when navigating quickly between the test and edit pages, top-level immutable fields are greyed out even when a SaaS config references no collection fields, and clicking a field whose YAML name is quoted (for example on or 1099) scrolls the editor to it.

Lethe

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

• Abuse Risk Scoring For Public Submissions: Publicly submitted privacy requests are scored for abuse signals — email local-part entropy, canonical duplicates, per-IP and per-email submission velocity via short-TTL Redis counters, and disposable or blocklisted email domains with a customer-configurable blocklist and allowlist. Scoring is configurable under the new security.risk_scoring settings. Includes database migrations.
• Flagged-For-Review Queue: With risk scoring enforcement enabled, requests crossing the flag threshold are held for manual review instead of processing automatically. The Request Manager gains a Flagged tab listing them, with status-gated approve and soft-delete actions and copy that reflects the current mode. Approving, denying, or cancelling a request clears its flagged state and returns it to the main list, retaining the score and signals. Includes a database migration.
• Hybrid SaaS Integrations: Installing a connector template that declares a manual task config now atomically creates a locked human-review manual task alongside the connection, surfacing the automated connector's discovered data to a reviewer as read-only context. Supports access review and erasure confirmation, with the bundled task locked against operator edits. Integration cards show "Manual tasks" and "API" capability tags, and hybrid connectors gain a manual-tasks tab and a reviewer-assignment setup step. Includes a database migration.
• Video Attachments On Manual Tasks: Manual task attachments accept video files (mp4, mov, avi). A submission containing any invalid or failed file is now refused atomically with an error naming the offending file, instead of completing the task with missing evidence.
• Wider Deadline Extension Caps: Response deadline extensions are capped for every modeled regulation with a statutory limit — UK GDPR (60 days), PIPEDA (30 days), and the US state privacy acts (45 days each, Florida FDBR 15 days) — matching the existing CCPA and GDPR budgets. Quebec Law 25 and LGPD extensions are blocked entirely, as neither permits a reviewer-discretion extension.
• DROP Requests Via API: DROP-sourced privacy requests can be created through the API under a dedicated privacy-request:create-drop scope, with idempotent replay scoped to the creating client. The DROP tab appears once DROP requests exist, and the recent-cycles table adds cycle duration, a live progress indicator for running cycles, a visible row-expand affordance, and persistent page size. Includes a database migration.
• Request Manager Filters: The Source filter gains a "No source" option for requests submitted without one.
• Bug Fixes:
  • Privacy request CSV export no longer times out on large datasets — the response is streamed.
  • The external manual task portal shows each task's actual request type instead of always showing "Access".
  • A wrong or expired verification code no longer kicks data subjects out of the DSR flow.
  • Manual task user assignment surfaces the specific permission error and clears the rejected user from the field when assignment fails.
  • Jira ticketing manual tasks are hidden from the Manual tasks tab.
  • Denying a pending request from the Request Manager list view no longer crashes the row.
  • The traversal preview's policy picker is scoped to policies bound to the selected property's privacy-center actions, so choosing one no longer returns "Policy not bound to this property".
  • The DROP archive-build retry window was extended so cycles no longer fail prematurely.
  • The external task portal in the Privacy Center is no longer unstyled.
  • The interrupted-task watchdog, exited-task poller, and discovery-monitor watchdog check Celery task liveness for their whole working set in one worker broadcast, so their duration no longer scales with the number of in-flight requests or monitor tasks.

Janus

Consent management and enforcement across web, mobile, and server-side surfaces

• Check Cookie Consent From FidesJS: A new Fides.getCookieConsent() API reports whether a given cookie is permitted under the user's current consent. Wildcard-aware, so patterns like _ga_-id- resolve correctly.
• IAB TCF Policy v5.0.b: Initial support for IAB TCF Policy v5.0.b and GVL specification version 4, including feature standard explanation texts and illustrations.
• Bidirectional Consent Integration Mapping: Privacy notices show which bidirectional consent integrations they are mapped into and warn when a mapped notice cannot propagate preferences. The notice list and detail APIs report each notice's integrations and enforcement level.
• Tunable Consent Metrics Rollups: Consent metrics rollup chunk sizing is exposed as configurable settings.
• Privacy Center Telemetry: Privacy Center server logs can be shipped to Ethyca's telemetry pipeline, opt-in via FIDES_PRIVACY_CENTER__TELEMETRY_PROTOCOL=opentelemetry.
• Bug Fixes:
  • The TCF consent modal no longer crashes when a feature or purpose has no embedded vendors or systems.
  • Legacy email identities are normalized so case-variant duplicate consent reads resolve correctly. Includes a database migration.
  • Setting consent metrics and translation settings through /api/v1/config no longer returns a 422.

Astralis

AI-assisted privacy assessments, access policies, and purpose-based access control

• Assessments v2 (Alpha): An alphaAssessmentsV2 feature flag gates the outcome-based assessments experience, including review workflows and Astralis coverage evaluations.
• Draft Access Policies From Documents: Backend support for generating draft access policies from uploaded documents and policy URLs. Includes a database migration.
• Astralis Dashboard (Alpha): An alphaAstralisDashboard feature flag replaces the Home page with the new Astralis dashboard experience.

Integrations

• Microsoft Purview: Available as an integration for datastore discovery — see Helios above.
• Bug Fixes:
  • Zendesk erasure is ordered so a user's tickets are deleted before the user, and an unclosed-ticket refusal now surfaces a clear message instead of a raw 422.
  • Mailchimp erasure DSRs no longer fail with a 400 on the member update — server-owned fields (ip_signup, unique_email_id, web_id, location country) are marked read-only so they are excluded from erasure while remaining visible in access.
  • The "Linked system" tab is restored on manual task integrations, and system links persist for every connection in a bulk PATCH rather than being dropped for pre-existing connections.

Fides core

• Backend-Driven Feature Flags: The backend feature-flag catalog is now the single declaration of which flags exist. Admin UI flag keys are generated from it through the OpenAPI schema, and the frontend flags.json has been removed.
• Multi-Org Request Context: The caller's active organization and accessible-organization set are carried on the request context, populated at the RBAC authorization chokepoint.
• Request Tracing: An OpenTelemetry span can be emitted per HTTP request, opt-in via FIDESPLUS__TELEMETRY__TRACE_REQUESTS=true.
• Postgres 17: The Fides application database moves from Postgres 16 to 17 for local development, tests, and CI.
• Security:
  • A policy engine sidecar data race that crashed the process when a cache invalidation landed during evaluation is fixed. Running the sidecar without token validation now requires an explicit DEV_MODE opt-in rather than being inferred from a missing encryption key.
  • pyasn1 bumped to 0.6.4, fixing two DoS advisories (CVE-2026-59885, CVE-2026-59886).
  • A 3-day minimum release age for npm packages reduces supply chain attack exposure.
  • DROP idempotent replay is scoped to the creating client, preventing cross-client access. Includes a database migration.
• Bug Fixes:
  • Feature-flag changes no longer appear stuck on deployments with a read replica — the flag store reads from the primary, so the cache always reflects the latest write.
  • The respondent role is granted feature_flag:read, so flag-gated UI resolves for respondents instead of always reading as off. Includes a database migration.
  • Privacy notices list loading no longer issues one query per notice.
  • The latest Docker image tag is no longer moved by a patch release cut on a superseded version line.
  • The full-fat Docker image builds again — it previously failed while pre-downloading the universal-sentence-encoder model, because tensorflow_hub imports the pkg_resources module that setuptools 83 removed. Affected versions published a slim image and then failed before the non-slim push, leaving releases that could not be deployed.
  • A performance index was added on custom_privacy_request_field.privacy_request_id to keep large privacy request CSV exports and cascade deletes fast. Includes a database migration.

Database schema & data changes

• This release includes database migrations for the flagged-for-review state and risk scoring signals, cloud infrastructure staged-resource error tracking, document-based access policy generation, Microsoft Purview discovery, hybrid SaaS integrations' bundled manual tasks, DROP idempotency client scoping, legacy email identity normalization, the respondent feature-flag read permission, and a custom_privacy_request_field.privacy_request_id index.

Notes

• The accessPackages and downloadAccessRequestResults feature flags are gone. Access packages and the access-results download are now controlled solely by their deploy-time settings. Deployments toggling these as feature flags need to set the corresponding settings instead.
• Postgres 17 for local development. Local dev, test, and CI stacks move to Postgres 17. Existing local databases built on 16 need to be recreated.
• Batch classification defaults changed. Batch-mode LLM classification now defaults to a 64-thread pool and 25 fields per call. Per-field classification keeps its existing 8-thread pool.