Fides 2.91.X release notes

• 2.91.1: August 5, 2026
• 2.91.0: August 3, 2026

Fides 2.91.1

Helios

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

• Restored the Legacy ML Classifier: The ML classifier dependencies (Presidio/spaCy) are restored in the full-fat Docker image. They were installed and then pruned again by every downstream image stage, so the legacy classifier could not run in any published image from 2.87.0 onward and its resources landed in a classification-error state.

Fides 2.91.0

Helios

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

• Batch Classification In The Monitor UI: The per-monitor batch-mode LLM classification setting is now exposed in the datastore discovery monitor configuration form. Batch mode defaults to a 64-thread pool and 25 fields per call, the configuration validated in benchmarking; per-field classification keeps its existing 8-thread pool.
• Faster Bulk Classification: Bulk-select classification in data discovery is coalesced into a single task, so classifying many resources at once runs noticeably faster.
• Classification Routes Around Degraded Providers: Batch-mode LLM classification probes upstream providers once per job and routes around any currently flagged as degraded, and operators can layer their own routing preferences via classify_params. Fail-open on any probe error.
• Recover Fields Stuck In Classifying: A new reset bulk field action moves fields stranded in a classifying state to classification error, so they can be classified, labeled, or ignored again.
• Action Center Improvements: Cloud infrastructure monitors gain promote-removal support, and approved asset visibility in the datastore monitor was reworked for clarity.
• Bug Fixes:
  • The infra monitor restore action regression was fixed, and AWS resource approval logic now behaves correctly.
  • The datamap report taxonomy filter dropdowns no longer re-render unnecessarily.
  • Stale monitor config key-mapping and privacy experience rows are cleaned out of the database cache. Includes a database migration.

Lethe

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

• Response Deadline Extensions: A privacy request's response deadline can now be extended by a set number of days with a required reason, with regime-aware caps (CCPA 45 days, GDPR 60 days), an audit log entry per extension, and an optional notification to the data subject. Available from the Admin UI request view. Includes a database migration.
• Access Package Review At Scale: The access package review experience was rebuilt around a prewarmed review artifact. Review pages load in milliseconds instead of re-streaming the whole results blob, the review tab loads lazily through a bounded skeleton and per-category endpoints, redaction edits validate against the artifact's field-path index rather than re-reading the full blob, and oversized packages export as NDJSON. Memory and CPU are bounded at the 50,000-record ceiling, and the previous redaction size limit is gone.
• Search Inside An Access Package: The access package review tab gains a search box covering field names and values across the whole package, served from the prewarmed artifact with windowed results and a retryable warming-up state.
• Hide The Source Column: Operators can hide the source column in access packages. With source hidden, duplicate values arriving from different sources collapse into a single reviewable row.
• Manual Tasks Without Attachments: A DSR manual task can be completed without an attachment, with an optional comment that is surfaced to the data subject in the access package.
• Request Workflows Is Now Generally Available: The Request Workflows (DSR traversal visualizer) page is out of feature flag. The preview also gained policy selection with per-policy cache keys, manual-task gating conditions, and manual-task assignees.
• DSR Policies And Form Builder Are Now Generally Available: DSR Policies and the privacy center form builder are no longer behind feature flags.
• Smarter Duplicate Detection: Duplicate detection now canonicalizes email identities so bot-submitted permutations of the same mailbox group together. Case is normalized for all domains; +tag subaddressing is stripped for providers that implement it (Google, Microsoft consumer, Proton); dots are additionally ignored for consumer Gmail only, and googlemail.com folds into gmail.com. Domains off the allowlist normalize case only. Includes a database migration.
• Skipped Erasure Visibility: Privacy requests that complete with skipped erasure tasks, for example because a connection was disabled, are now surfaced rather than silently completing: audit log and activity timeline entries, an email alert to configured DSR notification recipients, a "Skipped Erasure Datasets" column in the CSV export, and a warning banner in the request detail page.
• Submission Provenance: Publicly submitted privacy requests now capture client IP and user agent for abuse forensics, exposed to admins only. Includes a database migration.
• Pause Visibility In The Event Log: Execution log entries are recorded when a privacy request is paused by a webhook or a batch email send, visible in the Admin UI event log.
• Bug Fixes:
  • The DSR watchdog no longer errors privacy requests awaiting their first dispatch. Previously this intermittently canceled freshly-approved requests and emailed the subject an error, even though the request went on to process every task successfully.
  • Denying a pending request from the Request Manager list view no longer crashes the row.
  • Legacy dataset test requests no longer leak into the privacy request list. Includes a data migration that normalizes legacy source values.
  • Concurrent consent tasks can no longer propagate different preference states.
  • DSR graph construction is no longer O(N²) in identity-bearing fields, which speeds up graph builds on large datasets.
  • Manual-task config-field and connection edits are reflected in the traversal preview cache, and the Request workflows URL no longer carries an inert action-type segment.

Janus

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

• Consent Metrics Dashboard: The consent report page gains a metrics dashboard (behind a feature flag) showing opt-ins and opt-outs over time, totals, and breakdowns, backed by new rollup tables, a scheduled aggregation task, and a read API. Historical backfill is included so the dashboard covers past consent events, configurable via consent_metrics_backfill_days (default 365) and extendable later by raising the setting. Includes database migrations.
• Bug Fixes:
  • Privacy Center tiles now wrap onto multiple rows instead of overflowing when too many are configured to fit on one line.
  • The Privacy Center config description field is now optional.

Astralis

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

• Draft Access Policies From Documents: Access policies can now be drafted from an uploaded document or a policy URL, with Astralis generating the draft (alpha, feature-flagged). Includes a database migration.
• Faster Privacy Assessments: Privacy-assessment generation now runs its LLM calls through a bounded thread pool (FIDESPLUS__LLM__ASSESSMENT_LLM_CONCURRENCY, default 4), cutting generation time roughly in line with concurrency for identical output. Failed units now surface as partial completion instead of reporting a false 100%.
• Access Policy Builder Cleanup: The non-functional access policy onboarding page was replaced with a list empty state.

Integrations

• Greenhouse Harvest V3 Connector: A new greenhouse_v3 SaaS connector ahead of the v1/v2 API sunset, with OAuth2 client-credentials auth, an introspection-based connection test, candidate access traversal including EEOC and demographic data, inline option-id decoding, and anonymize-based erasure.
• Jira Ticket Templating: Data-subject custom privacy-request fields can be rendered into Jira ticket summaries and descriptions via per-field __CUSTOM_<KEY>__ template variables, and those fields now appear in the template-variable picker sourced from the Privacy Center config.
• Iterate: The users collection reports total_surveys and total_responses, and was renamed to survey_user.
• Bug Fixes:
  • Jira ticket rendering no longer re-substitutes template tokens that appear inside request values.

Fides core

• Property Branding: Properties gain structured branding fields (logo, colors, header, footer), with per-property access-package branding configurable in the Admin UI property form. Includes a database migration.
• Backend-Driven Feature Flags: Feature flags are now resolved from the backend flag service and applied per deployment; the legacy per-browser override layer backed by localStorage has been removed. The feature flags list on About Fides gains search, plus tags and tooltips covering whether a flag was set by an admin or by this deployment's config, when it was last changed, and what a reset restores.
• Configurable Invite Code TTL: The emailed user-invite code TTL is configurable, with the default raised to 7 days.
• Consistent Admin UI Forms: Admin UI forms were standardized onto a consistent shape with Fieldset sections, a shared form footer, save gating, and dirty-close guards.
• Security:
  • Query parameters are stripped from URLs in SaaS connector logs in production, and preserved in dev mode for debugging.
  • pillow bumped to 12.3.0 (four High-severity DoS CVEs and one shell injection CVE), httplib2 to 0.32.0 (High-severity decompression DoS CVE), and temporalio to ~=1.30.0 (four Rust crate advisories in its precompiled native bridge).
• Bug Fixes:
  • Approvers can now read feature flags, so flag-gated UI resolves for them instead of always reading as off. Includes a database migration.

Database schema & data changes

• This release includes database migrations for privacy request deadline extensions, document-based access policy generation, consent metrics rollup tables and their historical backfill, duplicate-detection email canonicalization, submission provenance capture, property branding fields, legacy dataset-test source normalization, database cache cleanup, and the approver feature-flag read permission.

Notes

• Database credentials are now required. The application requires a database user and password unless credentials come from a dynamic secrets provider. Deployments relying on implicit or empty credentials must set them explicitly before upgrading.
• Per-browser feature flag overrides are gone. Flags that were previously toggled per browser via localStorage are now resolved from the backend per deployment. Any workflow depending on a browser-local override needs to set the flag at the deployment level instead.