Highlights

Security advisory

The 9.5.5 release contains fixes for potential security vulnerabilities. Please see our security advisory for more details.

Features and enhancements

Aggregations

• Limit the multi_terms aggregation to a maximum of 128 fields [#157906]

Authorization

• Add the manage, create_index, read, index, write, and delete privileges on third-party agent indices to the kibana_system role [#160185]

ES|QL

• Upgrade the Parquet Hadoop bundle dependency [#158981]
• Upgrade zstd-jni to 1.5.7-16 [#159621]

Inference

• Pass the reasoning effort setting through to OpenAI chat completion requests [#156505]

Infra/Core

• Upgrade Jackson to 2.21.7 [#160501]

Network

• Upgrade Netty to 4.1.138.Final [#159561]

Fixes

CRUD

• Fix replica divergence on indices that have sequence numbers disabled [#158870] (issue: [#150408])

Distributed

• Prevent the allocation explain API from starting gateway shard data fetches [#157766] (issue: [#155449])

ES|QL

• Add circuit breaker accounting to the Base64 functions, MV_CONCAT, and MV_ZIP [#159659]
• Clear Parquet and ORC filter state between query reuse cycles [#158155]
• Disallow full text functions after UNION ALL when one of its branches contains a top N operation [#158578]
• Fix a ClassCastException in Parquet late-materialization filters when a predicate batch contains only null values [#157587] (issue: [#157313])
• Fix METADATA behavior with subqueries and logical views [#157233] (issue: [#160116])
• Fix a NullPointerException when folding a regular expression [#158768] (issue: [#158748])
• Fix TO_STRING being ignored in MATCH and MATCH_PHRASE on indexed fields [#159416] (issue: [#159395])
• Fix TO_TEXT being ignored in MATCH and MATCH_PHRASE on indexed fields [#159269] (issue: [#159265])
• Fix handling of the isPartial flag [#157475] (issue: [#157440])
• Fix an analyzer loop caused by repeated conversions [#159257]
• Fix nested fields conflicting with object fields [#158573] (issue: [#154011])
• Read partition_detection and partition_path on the query path [#157208]
• Reject TRANGE queries that have a null start or offset [#159125]
• Return a warning instead of failing on cumulative T-Digests [#160534]
• Return a warning when a GROK typed conversion fails [#160574] (issue: [#160564])

Engine

• Handle refresh exceptions in ensureShardSearchActive [#159526]

Extract&Transform

• Limit the length of connector descriptions [#159638]

Geo

• Limit geometry nesting depth when a geometry is constructed [#160004]
• Fix edge selection in the H3 great-circle latitude correction [#160228] (issue: [#159682])

Inference

• Convert RecursiveChunker to an iterative implementation and cap the size of its separator list [#158589]

Infra/Metrics

• Clear buffer files at startup [#159701]

Ingest

• Allow multipage PDFs in the semantic field [#159641] (issue: [#158773])

Machine Learning

• Detect level changes that coincide with the detection of new seasonality [#3209] (issue: [#2730])
• Stabilize uncertain trend forecasts [#3189] (issue: [#2772])

Mapping

• Resolve high-cardinality keyword queries to ScanningBinaryDocValuesAutomatonQuery [#158603] (issue: [#158597])

SQL

• Fix a NullPointerException and ClassCastException in ReplaceSumWithStats when SUM functions are nested [#159256] (issue: [#159250])

Search

• Fix an IllegalArgumentException (unknown transport version) in point in time (PIT) requests during mixed-version cross-cluster search [#159248] (issue: [#158991])
• Fix a lost listener in the query phase when a remote reduction is rejected [#160152] (issue: [#158625])
• Fix a request circuit breaker leak in the rank feature phase [#158951]
• Fix sorted search failures on empty shard results [#160132]
• Improve consistency checks for point in time (PIT) IDs [#160179]
• Release top_hits hits when a cross-cluster search merge reduce fails [#160363]
• Release late shard results after a phase failure [#160439] (issue: [#160321])
• Tolerate missing remote indices during the inference rewrite [#160207] (issue: [#160203])

Security

• Improve consistency checks for shard-level requests [#159691]

TSDB

• Reject Prometheus remote write 2.0 requests [#159987]

Vector Search

• Honor the precondition in the auto calibration error model [#158569]
• Use the effective manifold dimensionality when calculating the error standard deviation in auto calibration [#158158]

Highlights

Security advisory

The 9.4.8 release contains fixes for potential security vulnerabilities. Please see our security advisory for more details.

Features and enhancements

Aggregations:

• Limits the multi_terms aggregation to a maximum of 128 fields [#157906]

Authorization:

• Grants the kibana_system role the manage, create_index, read, index, write, and delete privileges on the XM Cyber indices (logs-xm_cyber.device-*, logs-xm_cyber.product-*, and logs-xm_cyber.vulnerability_index-*), so the index lifecycle management (ILM) delete step can succeed for these indices [#160185]

ES|QL:

• Upgrades the Parquet Hadoop bundle [#158981]
• Upgrades OpenTelemetry to 1.62.0 in the Google Cloud Storage (GCS) modules [#152499]
• Upgrades zstd-jni to 1.5.7-16 [#159621]

Inference:

• Passes the reasoning effort setting through to OpenAI chat completion requests [#156505]

Infra/Core:

• Upgrades Jackson to 2.21.7 [#160501]

Network:

• Upgrades Netty to 4.1.138.Final [#159561]

Fixes

CRUD:

• Fixes replica divergence on indices that have sequence numbers disabled [#158870] (issue: [#150408])

Distributed:

• Prevents the allocation explain API from starting gateway shard data fetches [#157766] (issue: [#155449])

ES|QL:

• Fixes a NullPointerException that occurred when ES|QL folded a regular expression that evaluated to null [#158768] (issue: [#158748])
• Fixes a failure (Rule execution limit [100] reached) when a query that reads from multiple sources, such as FROM (FROM a), (FROM b), applies the same conversion function to a field more than once [#159257]
• Fixes ES|QL queries that fail on indices where a nested field conflicts with an object field [#158573] (issue: [#154011])
• Makes the ES|QL TO_BASE64, FROM_BASE64, MV_ZIP, and MV_CONCAT functions respect the circuit breaker, so large inputs return a circuit breaker error instead of exhausting memory [#159659]
• Rejects TRANGE queries that have a null start or offset [#159125]
• Returns null values and a warning, rather than failing the query, when a typed GROK capture such as %{NUMBER:n:int} matches a value that can't be converted to the requested type [#160574]

Engine:

• Prevents spurious ERROR log entries when a shard closes while a refresh triggered by a search on a search-idle shard is still queued [#159526]

Extract&Transform:

• Limits the length of a connector description to 8192 characters [#159638]

Geo:

• Fixes incorrect H3 cell bounds. The latitude correction now checks both edges next to the extreme-latitude vertex instead of only one [#160228] (issue: [#159682])
• Limits the nesting depth of geometry collections when a geometry is created, preventing excessive recursion when a geo_shape runtime field emits deeply nested geometries [#160004]

Inference:

• Makes recursive chunking use constant stack depth regardless of input size, and limits the number of separators it accepts [#158589]

Infra/Core:

• Ignores unrelated files when scanning the configuration directory for log4j2.properties [#150016]

Machine Learning:

• Adds the max_consecutive_extraction_failures datafeed setting. A real-time datafeed now stops automatically after repeated consecutive extraction failures instead of retrying indefinitely [#159554] (issue: [#158426])
• Detects level changes that coincide with the detection of new seasonality [#3209] (issue: [#2730])
• Prevents anomaly detection job reopen from exhausting search scroll contexts during mass reassignment [#154925] (issue: [#153260])
• Stabilizes forecasts for time series with an uncertain trend [#3189] (issue: [#2772])

SQL:

• Fixes a NullPointerException and ClassCastException when an aggregate function is nested inside another aggregate, for example HAVING SUM(s) > 10 where s is an alias of SUM(int) [#159256] (issue: [#159250])

Search:

• Fixes a memory leak where shard results that arrived after a search phase failed were never released [#160439] (issue: [#160321])
• Fixes a request circuit breaker leak in the rank feature phase [#158951]
• Fixes an IllegalArgumentException (unknown transport version) when using a point in time (PIT) in cross-cluster search on mixed-version clusters [#159248] (issue: [#158991])
• Fixes failures in sorted searches when a shard returns an empty result after timing out [#160132]
• Adds validation that the shards in a point in time (PIT) ID match their reader contexts, so a corrupted PIT ID no longer returns unreliable search results [#160179]
• Releases top_hits results when the cross-cluster search merge reduce fails [#160363]
• Fixes searches that return a 404 error when a remote cluster is missing the queried index during the inference rewrite and ccs_minimize_roundtrips is false or a scroll is used [#160207] (issue: [#160203])

Security:

• Adds a check that shard IDs in cross-cluster requests that target specific shards match the current cluster state [#159691]

TSDB:

• Rejects Prometheus remote write 2.0 requests with an HTTP 415 response so senders can fall back to 1.0, instead of returning 204 and silently dropping the samples [#159987]

Fixes

Elastic Agent Builder:

• Fix Agent Builder tools hanging or timing out on data streams that span the frozen tier. Tools now skip frozen tier indices #292349.
• Fix Agent Builder tracing applying the default space's GenAI privacy settings in every space. Each conversation now uses the privacy settings of its own space #290750.
• Add the case template ID and version fields to the Agent Builder cases-analytics skill so agents can filter and group cases by template #291716.

Alerting and cases:

• Fix forced rule runs and schedule changes being silently dropped when you request them while the rule is finishing a run #293492.
• Fix Agent Builder rules failing after about an hour because they kept the agent's temporary API key #291318.
• Fix bulk alert tag updates also changing matching alerts in other Kibana spaces #285874.
• Fix Jira Service Management and Opsgenie connector actions failing when the rendered alert message exceeds 130 characters. The connectors now truncate the message instead #291698.
• Fix the email connector HTTP test sending caller-supplied subject and message content. The test now sends fixed content #288664.
• Fix the Auto-push case to connected external system option missing from the Cases rule action when case templates are enabled #292187.

Dashboards and Visualizations:

• Increase the TinyMath expression length limit so longer visualization formulas don't fail #290314.
• Fix Go to URL drilldowns leaving Mustache tokens such as event.values.[0] unreplaced in the drilldown name #290363.
• Fix ES|QL metric chart background trendlines failing for queries that use FORK #287819.
• Fix new Last value columns being exported with Show array values turned on even when the option is off, which caused Elasticsearch errors when you imported them #292864.
• Fix filter pills showing filter value is invalid or incomplete when the filter value is an empty string #290383.
• Fix clicking the unsaved changes indicator on the dashboard save button opening two save dialogs #293087.

Data ingestion and Fleet:

• Fix deleting a managed integration policy leaving the agent listed as active #290181.
• Fix agents for Elastic managed integrations enrolled in other Kibana spaces appearing in the Fleet Agents list #293142.
• Fix version-specific agent policy assignment failing repeatedly for agent policies in non-default Kibana spaces #293010.
• Fix the Fleet output edit flyout freezing or showing no counts when thousands of agent policies use the output #291972.
• Fix Elastic Agent upgrade and agent policy requests hanging for up to two minutes and failing with a reset connection when Kibana can't retrieve the list of available Elastic Agent versions #292788.
• Fix integration upgrades failing when you add a custom analyzer or normalizer to a @custom component template after the write index is created. Fleet now rolls over the data stream instead #291813.
• Fix Fleet installing only the first trained model when an integration package includes several #290145.
• Fix ambiguous_conflict errors on dashboard assets when you install an integration in multiple Kibana spaces #291648.
• Fix installed integration icons rendering as black shapes #293176.
• Fix the Edit pipeline button in the ingest pipeline details flyout moving off-screen when a processor contains a long value #291000.

Data management:

• Fix Index Management moving component templates listed in ignore_missing_component_templates to the end of composed_of when you edit or clone an index template #290171.

Developer tools:

• Fix Console reporting Elasticsearch request timeouts as connection failures. Console now explains that the request might still be running and suggests checking its status before you retry #289481.

Elasticsearch solution:

• Add an Expand site group members toggle to SharePoint Online connectors for compact document-level security #288070.
• Add a Sync all mail folders option to the Outlook connector #291396.
• Fix deleted content connectors remaining in the connectors list #290859.

Elastic Observability solution:

For the Elastic Observability 9.5.5 release information, refer to Elastic Observability Solution Release Notes.

Elastic Security solution:

For the Elastic Security 9.5.5 release information, refer to Elastic Security Solution Release Notes.

Kibana platform:

• Hide the Activate user action on deactivated users when you have read-only access #289801.
• Fix a failed audit log write after startup crashing Kibana. Kibana now reports a degraded state and turns the audit logger off instead #287360.
• Fix Kibana shutting down when the product documentation installer downloads a corrupt or incomplete artifact #291142.
• Fix PDF and PNG reports using too much memory and failing on memory-constrained instances when you generate them from a wide browser window #289062.

Machine learning and inference:

• Fix streaming inference requests returning no content when elasticsearch.compression is enabled, which made Agent Builder report an empty response #290381.
• Fix LLM feature settings listing every connector when you lack inference privileges, instead of only the allowed connectors #290457.
• Fix LLM resolution failing when you lack the Actions privilege and an inference endpoint is configured in model settings #290533.
• Fix the anomaly detection bucket span estimation API accepting an unlimited number of detectors, which could exhaust Kibana server resources #291203.
• Fix the field statistics flyout in the anomaly detection job wizard closing on its own while the wizard validates the job #292297.

Workflows:

• Fix the elasticsearch.indices.exists workflow step crashing the workflow or returning an output you can't use in conditions #289580.

Elastic today announced the general availability of cross-project search (CPS) for Elastic Cloud Serverless, enabling organizations to run a single query across multiple serverless projects.

With CPS, customers can search across all regions, project types, and cloud providers, with easy setup, versionless compatibility, and complete security scope. Data is queried in place, eliminating data movement risk and keeping egress costs minimal. It supports all solutions, including Observability, Security, Search, and vector database workloads, giving teams centralized visibility without the operational overhead of centralizing data.

Organizations manage data across projects for legitimate reasons: data residency compliance, tenant data isolation, or organizational structure needs. But it can create a painful side effect: no unified visibility. SOC teams jump between regional projects chasing a threat. Site reliability engineers (SREs) can't see production and staging signals in one place. The workaround, centralizing data, means storing and moving the same bytes twice, and risking sovereignty violations. Existing solutions don't help either; some are locked to a single cloud vendor, while others require version coordination and server setup just to run a federated query.

CPS truly breaks down data silos across distributed architectures without compromising data isolation, residency, or boundaries that systems rely on. By delivering transparently merged results in a single query, CPS allows teams to analyze cross-departmental security signals side by side or seamlessly correlate production and staging logs, or developing search apps with context from multiple linked projects, all without forcing separate deployments into a single environment.

Cross-cluster search (CCS), a similar feature in Elastic Cloud Hosted (ECH), a fully managed cloud service for running the Elastic Stack, is already used by more than 80% of large ECH customers with complex architectures, some managing over 50 remote cluster connections on a single deployment. CPS delivers the same capability on Serverless without the need to understand deployment topology, configure certificates, or per-connection authentication. CPS infrastructure is abstracted by design: three clicks in the Cloud console to link projects, with authentication handled at the org level by Elastic's control plane.

That is the entire setup needed, period.

Once linked, searches in Discover, ES|QL, dashboards, and alerting rules run across all linked projects automatically, with the Kibana CPS Scope Picker keeping scope firmly under your control.

CPS: The Serverless-native approach

Here’s how it works: From your project settings, select the projects you want to link. Filters and tags help target the right environments. Once saved, searches in Discover, dashboards, and alerting automatically include data from every linked project.

Simple setup:

Linked projects data is queried in Discover by default.

The CPS scope selector in Kibana lets analysts and engineers switch query scope without changing query syntax — filtered by project tags or just the current one.

Administrators set the default per Kibana space. A space used for local incident investigation restricts default scope to a single project; a global operations space defaults to the full linked set. Users can override the default for any individual query or session using the project picker, without touching the space configuration.

What’s new in general availability?

CPS launched in tech preview on April 16, 2026. This GA release ships with meaningful additions.

Scale: Support for up to 100 linked projects is included by default. Enterprise teams and MSSPs that link per-client or per-region projects can now operate at the scale their architectures require. Higher limits are available upon request.

Serverless project as an origin: At GA, existing projects can link and query across as many as 100 linked projects without requiring migration or recreation.

ML and AI capabilities: With CPS, Elastic ML jobs detect anomalies in time-series data unsupervised, and at GA they can query across multiple linked projects in a single job. Application performance monitoring (APM) anomaly detection in Elastic Observability can also span linked projects, giving SRE teams a unified anomaly view. Moreover, Elastic Agent Builder enables AI agents across all solutions to retrieve context from multiple linked serverless projects, establishing CPS as a foundational building block for production agents operating on distributed data.

Versionless and cloud-agnostic: Teams aren't locked into a single cloud vendor or forced to coordinate versions across deployments just to do a search. And it works across every solution: Observability, Security, Search, and vector database workloads.

Observability: Application and infrastructure monitoring, alerting, SLOs, and APM anomaly detection work across linked projects enabling flexible data locality and analysis.

Security: CPS powers the Central SOC model: Detection rules run once on a central project and query across every linked project, so SOC teams triage one alert without manually merging data or duplicating alert rules. Entity information and Elastic Defend telemetry from linked projects feed that same central view.

Search: ES|QL queries and Discover sessions span linked projects by default. Across all solutions at GA an AI agent can retrieve context from multiple linked projects in a single query, making CPS a foundational building block for production agents that operate across distributed data.

Complete access control: A user's access to data across linked projects is determined by their individual permissions on each project, not by where they're querying from. What that user can see is always complete and consistent with what they are authorized to see.

Reduced costs: Because data stays where it is, you pay for the queries you run, not for storing and moving the same bytes twice.

Full project routing: It gives teams precise control over which projects a query targets. Using predefined tags such as region, cloud provider, and project alias — or custom tags you define — you can route any query to a specific subset of linked projects. Named expressions and boolean logic (AND, OR, NOT) let you build reusable routing rules for common patterns, such as "all EU projects" or "all MSSP tenant environments except staging."

Programmatic CPS access: CPS has introduced a new user role that enables anyone, not just organization owners, to generate Cloud API keys for CPS queries, removing friction for teams building pipelines and applications on CPS.

Frequently asked questions

What is CPS in Elastic Cloud Serverless? CPS lets you run a single query across multiple Elastic Cloud Serverless projects from Discover, dashboards, or alerting. Results are merged transparently. Setup takes minutes from your project settings.

How is CPS different from CCS? CCS is the equivalent feature in Elastic Cloud Hosted. CCS requires mapping your deployment architecture, managing API keys or certificates per remote connection, and referencing remote cluster names in queries. CPS handles authentication and topology automatically — link projects in the UI, and searches span all of them by default.

How much does CPS cost? CPS has two pricing dimensions. Mounted data is charged at $0.009 per GB retained in linked projects per month (pricing varies by project types, region and cloud service provider). Data egress out is charged at $0.05 per GB — the same rate as standard non-CPS egress, tracked on a separate line item so you can see exactly what CPS queries are contributing to your transfer costs. A team ingesting 10 GB per day with 30-day retention and one linked project adds approximately $4.50 per month in mounted data costs, plus egress based on query volume — typically a small fraction of total CPS cost for common single-pane-of-glass usage.

Which project types and tiers include CPS? CPS is available on all serverless project types. For Observability and Security projects, the Complete tier is required.

Can I control which projects are included in a search? Yes, at both the UI and query level. In Kibana, the project picker lets users switch between all linked projects and just the current one without changing query syntax; administrators set per-space defaults to control the scope for each team or workflow. At the query level, project routing parameters in ES|QL and the search API let you target specific linked projects directly, without going through the UI.

Get started

If you are already running Elastic Cloud Serverless, CPS is available in your project settings today. Link your projects, open Discover, and run your first cross-project query.

Not yet on Elastic Cloud Serverless? Sign up for free!

The release and timing of any features or functionality described in this post remain at Elastic's sole discretion. Any features or functionality not currently available may not be delivered on time or at all.

The 9.5.4 release contains fixes for potential security vulnerabilities. Check our security advisory for more details.

Fixes

Elastic Agent Builder

• Fix MCP tools dropping extra arguments that the schema allows through additionalProperties before execution #287815.

Alerting and cases

• Relabel the top-level Alerting privilege as Alerting V2 in the Roles UI and mark it experimental #288163.
• Fix the Tracking containment rule editor breaking when you select a data view #289778.
• Fix Sync alerts and Extract observables not applying when you change a case template from the sidebar #288636.
• Fix Alerting V2 rules from not saving when you remove all rule tags #288962.

Dashboards and Visualizations

• Fix the Update filter button being unreachable when the filter editor is taller than the screen #289163.
• Fix sibling dashboard panels not dimming during inline editing on dashboards that have no controls #288939.
• Fix ES|QL visualizations dropping field variable controls (??field) after you save them through the Visualizations API #288785.
• Fix Options list filters stopping after you delete a collapsible dashboard section #287923.

Data ingestion and Fleet

• Fix Fleet blocking a package policy save when every enrolled agent is incompatible with the package version, including agents that report a pre-release version. Fleet shows a warning instead of blocking the save #289596.
• Fix managed integrations appearing in the Fleet Agents list when they use a version-specific policy #289394.
• Fix package policy space assignments becoming stale after you move an agent policy to another Kibana space #288966.
• Fix Elastic Defend metadata transforms accumulating as duplicates in Elasticsearch across stack upgrades #288379 #288902.
• Fix comment-only values in Fleet agent policy advanced YAML settings leaving agents stuck on an outdated policy #288430.

Data management

• Fix classic streams leaving a stale Elasticsearch data stream mappings override after you remove all field overrides #288788.

Developer tools

• Fix Console treating request-like lines inside an unterminated triple-quoted string as standalone requests #287547.

Discover

• Cancel in-progress async searches in Discover and on dashboards when you navigate away in the browser, so Elasticsearch does not keep running them #287949.
• Fix custom time-range presets from the Time filter quick ranges advanced setting showing raw time values instead of their labels, and add a (rounded) suffix to rounded relative ranges #289066.

Elasticsearch solution

• Add an Expand role members toggle to ServiceNow connectors for compact document-level security #288071.
• Fix Query Rules document preview failing when the document ID contains slashes #287214.

Elastic Observability solution

For the Elastic Observability 9.5.4 release information, refer to Elastic Observability Solution Release Notes.

Elastic Security solution

For the Elastic Security 9.5.4 release information, refer to Elastic Security Solution Release Notes.

Kibana platform

• Fix unexpected logouts after an access token is refreshed during a search or dashboard refresh #286038.
• Fix user profile rows using the wrong height in user pickers #288451.

Machine learning and inference

• Fix Anomaly Explorer hanging when you add a geo job chart to a case #289363.

Features and enhancements

Alerting and cases

• Warn in the field library and template editors when a required case field has no default value, because automated case creation can't fill it in, and show a field's name and type as permanent as soon as you create it #285453.
• Enforce required global fields on API and workflow case creation when case templates are enabled, and accept extended_fields on the public and workflow case create and update APIs #283350.

Developer tools

• Raise the Dev Tools Console Copy to curl limit so a request-body string of up to 100000 characters copies successfully #287526.

Fixes

Elastic Agent Builder

• Fix the research agent re-asking clarifying questions you already answered in the same round #284800.
• Fix disabled pre-execution workflows disappearing from Agent Builder settings so you cannot remove them. A disabled workflow stays visible and is labeled (disabled) #286826.

Alerting and cases

• Fix legacy custom fields and their v2 field-definition counterparts falling out of sync, by linking each legacy field to its definition and keeping both representations equal on every write #282772.
• Fix cases created from a case template by a rule action not inheriting the template's external connector, Sync alerts and Extract observables settings, and assignees #284854.
• Fix matching fields missing from Cases Activity search results #284940.
• Fix case attachment UI issues, including delete controls on saved-object attachments and the more-actions menu for read-only access #285456.
• Fix saved Lens visualizations missing from case attachments, and add helper text in the markdown editor for inline Timeline and Lens references #286164.
• Fix creating a case from a workflow, a case template, or machine learning so Sync alerts is not turned on unless that solution enables it #286670.
• Fix the Additional fields editor on the ServiceNow ITSM, ServiceNow SecOps, and Jira connectors rejecting Mustache context variables as invalid JSON #286578.
• Fix required-on-close validation skipping global fields that a template includes by $ref #287306.
• Fix a template $ref global field ignoring its hide-on-value condition when the controlling field is in another section #287308.
• Fix global field column selections not syncing between the Cases list and table views #287327.
• Fix the Cases list keeping the previous page after you change filters #287945.
• Fix custom threshold no data alerts firing for groups that belong to a different custom threshold rule #287151.
• Fix custom threshold Rate aggregations not firing when a counter increases from 0 to a positive value #283973.
• Fix experimental alerting system .rule-events and .alert-actions data streams requesting a replica on single-node clusters, which left those shards unassigned #287544.
• Fix alerting resource installation using unbounded concurrency, which could exhaust Kibana heap when many contexts and namespaces install at once #287643.

Dashboards and Visualizations

• Fix by-reference visualization panels disappearing from dashboards when the panel's savedObjectRef reference is missing #285476.
• Fix metric chart background trendlines failing for ES|QL time-series queries that use TBUCKET(...) #284839.
• Fix the Maps timeslider close button registering clicks only on the top quarter of the control #287441.
• Fix Canvas functions API creating data tables with more than 50 columns per row #287935.

Data ingestion and Fleet

• Fix package policy inputs for newer agent versions not being backfilled after a Kibana minor upgrade #284710.
• Avoid querying Elasticsearch for integration sync when no cross-cluster replication follower index exists #285526.
• Fix space awareness on Fleet agent tag, upload, file, and unenroll API handlers #284936.
• Fix Fleet wrapping a Kafka Dynamic topic multi-field format string such as %{[data_stream.type]}-%{[data_stream.namespace]}, which produced an invalid topic expression #285581.
• Fix Fleet package manifests accepting cluster privileges outside the documented allow list #286084.
• Fix Fleet agent CSV export failing when xpack.reporting.csv.scroll.strategy is scroll, or when Search in frozen indices is enabled #286119.
• Fix the Fleet agent-status-change task failing when the Elasticsearch response is too large #286324.
• Fix sorting the Agents table by Host so it uses the hostname keyword field instead of the analyzed text field #286383.
• Fix package policies accepting a whitespace-only namespace, which saved while ingestion failed silently #286385.
• Fix the CPU and memory tooltip on offline agents showing misleading messages instead of clearly stating the agent is offline #286386.
• Fix managed integration throughput index patterns for streams that use a dynamic dataset #286758.
• Fix update and delete of legacy managed integration policies returning HTTP 404 or reporting success while the deployment kept running #287011.
• Fix stale version-specific Fleet policies remaining after their agent version dropped out of the bounded set, and refresh variants that still have enrolled agents #287095.
• Fix slow Fleet outputs fetching on Fleet setup and Integrations UI page load #287113.
• Fix Fleet API-key invalidation failing when more than 10,000 keys need to be invalidated at once #287252.
• Fix the Fleet integration policy editor showing the wrong credential method for policies created before the package introduced var_groups #287389.
• Fix the version-specific policy icon showing on agents that already meet the policy's version requirements #287504.
• Fix the Fleet policy editor not validating data_stream.dataset on integration packages #287557.
• Fix managed integration orphan cleanup so it respects package-policy ownership and the force flag #287571.
• Fix Fleet package policies keeping stale secret references after you rotate credentials #287642.

Developer tools

• Fix Dev Tools Console autocomplete not triggering at request-body completion positions, and fix template expansion and text corruption when accepting a suggestion mid-string #284530.
• Fix Dev Tools Console body autocomplete after a closed triple-quoted string value #286695.
• Fix Console autocomplete inserting boolean and numeric suggestions as quoted strings instead of JSON primitives #286699.
• Fix Console body autocomplete preferring a same-name global rule over an explicit field rule #286703.
• Fix duplicate Cut, Copy, and Paste actions in the Dev Tools Console context menu #286704.

Discover

• Fix ES|QL filters generating invalid casts such as ::text for types ES|QL does not support #287416.

Elastic Observability solution

For the Elastic Observability 9.5.3 release information, refer to Elastic Observability Solution Release Notes.

Elastic Security solution

For the Elastic Security 9.5.3 release information, refer to Elastic Security Solution Release Notes.

Kibana platform

• Fix the date range picker rejecting Elasticsearch date math that combines rounding with a further offset, such as now/y+3M. Chained expressions stay as shorthand instead of converting to absolute dates #287010.
• Fix audit log records omitting the client IP address for requests made over HTTP/2 #285344.
• Fix PKI authentication sessions being invalidated when in-flight HTTP/2 requests are cancelled, which logged you out. This affects deployments with server.protocol: http2, the default in 9.0.0 when TLS is enabled #285153.
• Fix Kibana failing to start when the file-based audit log appender cannot write. Kibana reports a degraded state instead #282347.

Machine learning and inference

• Fix hung LLM response streams blocking the Kibana event loop #286849.
• Fix chat inference requests using a non-default connector when genAiSettings:defaultAIConnectorOnly is enabled #288107.

Workflows

• Fix generated workflow edits stripping YAML comments and formatting #288075.
• Fix single-step workflow runs dropping variables from the historical context override #287347.
• Fix the Filter executions dropdown showing the same Waiting label for waiting-for-input and waiting-for-child statuses. The dropdown now uses Waiting, Waiting for input, and Waiting for child workflow #285345.

Features and enhancements

Authorization

• Grant kibana_system access to PingDirectory indices so ILM can delete them #156029

FIPS

• Upgrade Bouncy Castle FIPS to 1.0.2.7 and 2.0.2 #156296

Infra/Core

• Upgrade ASM to 9.10.1 #156213
• Upgrade the bundled JDK to 26.0.2 #156203

Machine Learning

• Add EuroBERT and Jina v5 ops to graph validation allowlist #3015
• Harden pytorch_inference with TorchScript model graph validation #3008 (issue: #2890)
• Improve adherence to memory limits for the bucket gatherer #2848
• Improve error handling for quantiles state documents #2894
• Improve handling of invalid JSON state documents #2895
• Improve messages when a process is stopped after running out of memory #2841
• Lower the log severity of recoverable errors #2889
• Report the actual memory usage of the autodetect process #2846
• Restrict file system access for PyTorch models #2851
• Upgrade PyTorch to 2.7.1 #2863

Search

• Improve search throughput by removing a lock in query cache stats lookup #156902

Security

• Upgrade httpclient5 to 5.6.4 #157564

Fixes

Aggregations

• Fix percentiles aggregations failing with an already closed breaker error #157254 (issue: #157206)
• Fix rank estimation ignoring min and max in exponential histograms #156268 (issue: #156057)

CRUD

• Retry writes when the target primary is unassigned during a shard split #157255 (issue: #156982)

Distributed

• Fix deletes silently skipped on TSDB indices that use synthetic _id #156903 (issue: #151326)

ES|QL

• Account for deferred aggregation memory in the request circuit breaker #155600 (issue: #148516)
• Account for t-digest percentiles memory in the request circuit breaker #155586
• Fix TS queries that name a time bucket after the timestamp field #155623
• Fix incorrect results when sorting on multiple keys with a LIMIT #155923
• Fix unresolved columns in CASE with CIDR_MATCH and implicit ip casts #156162 (issue: #155979)
• Reduce overhead of TS queries by pruning unused dimension fields #157250
• Reject 2^63 when converting double to long #157555 (issue: #157551)
• Reject COUNT_DISTINCT precision thresholds outside the integer range #157553 (issue: #157549)

Inference

• Fix Google Vertex AI tool calls with non-string argument values #156665 (issue: #156644)
• Fix inference and security cache clearing on multi-node clusters #156603

Infra/Metrics

• Fix NullPointerException from async telemetry instruments at shutdown #157161

Machine Learning

• Fail closed when the TorchScript state hook scan is incomplete #3149
• Fail gracefully when restoring a categorizer with an out-of-range token ID #3143
• Fix LFU cache count drift when a lock times out #3090
• Fix flaky CIoManagerTest/testFileIoGood test #3017
• Fix machine learning daily maintenance failing when security is enabled #148554
• Mark the machine learning controller non-dumpable before accepting commands #3081
• Re-audit repeated datafeed extraction errors instead of reporting once #157565
• Reject TorchScript custom state hooks and inductor::_reinterpret_tensor #3078
• Reject non-native ABIs in the machine learning seccomp filter #3080

SLM

• Fix SLM stats recording the wrong outcome for concurrent snapshots #155934 (issue: #155621)

SQL

• Reject ROUND results that overflow byte or short #157554 (issue: #157550)

Search

• Ensure batched queries always return a response #155908
• Fail the request instead of the node on deeply nested aggregations #155745
• Fix nested and mode-based sorts on integer fields returning incorrect values on indices created before 8.19 or in 9.0.x #155580 (issue: #155243)
• Fix search failures when a segment closes during a query cache stats update #157050
• Keep scroll contexts on transient search queue rejection #155697 (issue: #90912)
• Limit aggregation nesting depth with search.aggs.max_nested_depth #155743
• Reject null values in search_after instead of failing the search #132434 (issue: #132370)
• Stop returning a user field named _type as search hit metadata #155706 (issue: #110438)

Snapshot/Restore

• Halt the node on fatal errors during S3 multipart uploads #156765

Transform

• Preserve security headers when copying a transform configuration #156191

Vector Search

• Fix search failures when many field patterns match no vector fields #156466
• Score nested kNN inner_hits like the query phase #156332 (issue: #138496)