CVE-2026-55178

GeoLens is a self-hosted geospatial data catalog with semantic search, OGC and STAC APIs, and a map builder. Prior to 1.2.3, multiple read and link endpoints authorize only the resource named in the request URL and fail to re-authorize a second caller-influenced dataset reached through a relationship, map layer, VRT source, externalId lookup, or request body. When a public map references a private dataset, anonymous callers can use GET /maps/{id} and GET /maps/{id}/style.json to obtain the private layer's metadata, sampled values, or vector tiles. The style response can expose a replayable HMAC tile URL that is bound to neither a user nor a map. When a public source dataset has a relationship to a private target dataset, anonymous callers can use the dataset relationship APIs to enumerate the relationship and read rows from the private target's backing table. Anonymous callers can also use GET /collections/datasets/items with an externalId dataset UUID to obtain metadata for any private, restricted, or unpublished dataset because that lookup performs no visibility check. Authenticated users with the default editor role can mosaic another user's private raster into an owned VRT and read its pixels, and POST /ai/metadata/{summary,keywords,lineage,quality-statement} accepts a body-controlled dataset_id without a visibility check and returns private metadata and sample values. Pre-existing vrt_source_links also expose unauthorized member metadata and health unless each member is filtered at read time. These paths can disclose vector geometries and attributes, raster pixels, table rows, table names, column schemas, feature counts, extents, source URLs and filenames, contacts, and sampled row values. This issue is fixed in version 1.2.3.

GeoLens is a self-hosted geospatial data catalog with semantic search, OGC and STAC APIs, and a map builder. Prior to 1.2.3, multiple read and link endpoints authorize only the resource named in the request URL and fail to re-authorize a second caller-influenced dataset reached through a relationship, map layer, VRT source, externalId lookup, or request body. When a public map references a private dataset, anonymous callers can use GET /maps/{id} and GET /maps/{id}/style.json to obtain the private layer's metadata, sampled values, or vector tiles. The style response can expose a replayable HMAC tile URL that is bound to neither a user nor a map. When a public source dataset has a relationship to a private target dataset, anonymous callers can use the dataset relationship APIs to enumerate the relationship and read rows from the private target's backing table. Anonymous callers can also use GET /collections/datasets/items with an externalId dataset UUID to obtain metadata for any private, restricted, or unpublished dataset because that lookup performs no visibility check. Authenticated users with the default editor role can mosaic another user's private raster into an owned VRT and read its pixels, and POST /ai/metadata/{summary,keywords,lineage,quality-statement} accepts a body-controlled dataset_id without a visibility check and returns private metadata and sample values. Pre-existing vrt_source_links also expose unauthorized member metadata and health unless each member is filtered at read time. These paths can disclose vector geometries and attributes, raster pixels, table rows, table names, column schemas, feature counts, extents, source URLs and filenames, contacts, and sampled row values. This issue is fixed in version 1.2.3.

Summary

Multiple GeoLens read/link endpoints authorized only the resource named in the request URL (a map, a VRT, a source dataset, an AI request) and failed to re-authorize a second, caller-influenced dataset that the request reached through a relationship, layer reference, mosaic source, or request body. This "authorize the URL resource, read a different dataset un-re-authorized" pattern let callers read data from datasets they have no access to.

The most severe instances require no authentication at all (anonymous, network-only). Others require only the default editor role that any self-service signup / upload user receives.

All issues are fixed in 1.2.3. There is no complete configuration workaround — upgrading is the only full remediation.

Impact

Depending on the endpoint, an attacker can read, for datasets they cannot otherwise access:

• the dataset's vector tile data (actual feature geometries/attributes),
• the dataset's raster pixels,
• backing-table rows,
• and metadata — table name, column schema, feature count, extent, source URL/filename, contacts, and sampled row values.

Affected versions

All versions prior to 1.2.3 (includes the published 1.0.0, 1.2.0, and 1.2.2 releases and their PyPI/npm/GHCR artifacts). Fixed in 1.2.3.

Findings

1. Anonymous metadata + private vector-tile disclosure via public maps (PR #235) GET /maps/{id} and GET /maps/{id}/style.json authorized the map but not each layer's backing dataset. A public map that references a private dataset leaked that dataset's table name, column schema, feature count, extent, and sampled values to anonymous callers. style.json additionally returned a vector-tile URL carrying an HMAC signature bound to neither user nor map, which the tile endpoint accepts for non-public datasets with no user check — so the signature is replayable to read the private dataset's actual vector tiles. (Anonymous · High)

2. Anonymous private-row disclosure via dataset relationships (PR #234) The dataset FK-relationship APIs authorized only the source dataset from the URL, never the relationship target. A public dataset with a relationship to a private dataset let an anonymous caller enumerate the relationship (obtaining the private target's id/title and the relationship id) and then call the related-record endpoint to read rows from the private target's backing table. (Anonymous · High)

3. Anonymous metadata disclosure via OGC externalId lookup (PR #236) GET /collections/datasets/items?externalId=<uuid> resolved the dataset by id and returned the full OGC catalog record (title, summary, bbox, keywords, contacts, distributions, source org) with no visibility check — the user was never threaded into the lookup. An anonymous caller could read any private, restricted, or unpublished dataset's metadata by UUID. (Anonymous · High)

4. Cross-tenant raster pixel disclosure via VRT mosaics — SEC-C (PR #237) An authenticated user with the default editor upload permission could mosaic another user's private raster into a VRT they own, then read the victim's pixels back through raster tile / quicklook / COG endpoints that authorize only the attacker-owned VRT. VRT member pixels are compiled into one served asset and cannot be filtered at read time, so the fix authorizes every source dataset at write/link time. (Authenticated editor · High)

5. Cross-tenant metadata/sample-data disclosure via AI metadata endpoints — SEC-D (PR #238) The POST /ai/metadata/{summary,keywords,lineage,quality-statement} endpoints were gated only by the use_ai_chat permission (held by the default editor role). The attacker-controlled dataset_id in the request body flowed into the LLM prompt context with no visibility filter, rendering any dataset's title, summary, source URL, filename, column schema, and sample values into the response.…

GeoLens is a self-hosted geospatial data catalog with semantic search, OGC and STAC APIs, and a map builder. Prior to 1.2.3, multiple read and link endpoints authorize only the resource named in the request URL and fail to re-authorize a second caller-influenced dataset reached through a relationship, map layer, VRT source, externalId lookup, or request body. When a public map references a private dataset, anonymous callers can use GET /maps/{id} and GET /maps/{id}/style.json to obtain the private layer's metadata, sampled values, or vector tiles. The style response can expose a replayable HMAC tile URL that is bound to neither a user nor a map. When a public source dataset has a relationship to a private target dataset, anonymous callers can use the dataset relationship APIs to enumerate the relationship and read rows from the private target's backing table. Anonymous callers can also use GET /collections/datasets/items with an externalId dataset UUID to obtain metadata for any private, restricted, or unpublished dataset because that lookup performs no visibility check. Authenticated users with the default editor role can mosaic another user's private raster into an owned VRT and read its pixels, and POST /ai/metadata/{summary,keywords,lineage,quality-statement} accepts a body-controlled dataset_id without a visibility check and returns private metadata and sample values. Pre-existing vrt_source_links also expose unauthorized member metadata and health unless each member is filtered at read time. These paths can disclose vector geometries and attributes, raster pixels, table rows, table names, column schemas, feature counts, extents, source URLs and filenames, contacts, and sampled row values. This issue is fixed in version 1.2.3.

GeoLens is a self-hosted geospatial data catalog with semantic search, OGC and STAC APIs, and a map builder. Prior to 1.2.3, multiple read and link endpoints authorize only the resource named in the request URL and fail to re-authorize a second caller-influenced dataset reached through a relationship, map layer, VRT source, externalId lookup, or request body. When a public map references a private dataset, anonymous callers can use GET /maps/{id} and GET /maps/{id}/style.json to obtain the private layer's metadata, sampled values, or vector tiles. The style response can expose a replayable HMAC tile URL that is bound to neither a user nor a map. When a public source dataset has a relationship to a private target dataset, anonymous callers can use the dataset relationship APIs to enumerate the relationship and read rows from the private target's backing table. Anonymous callers can also use GET /collections/datasets/items with an externalId dataset UUID to obtain metadata for any private, restricted, or unpublished dataset because that lookup performs no visibility check. Authenticated users with the default editor role can mosaic another user's private raster into an owned VRT and read its pixels, and POST /ai/metadata/{summary,keywords,lineage,quality-statement} accepts a body-controlled dataset_id without a visibility check and returns private metadata and sample values. Pre-existing vrt_source_links also expose unauthorized member metadata and health unless each member is filtered at read time. These paths can disclose vector geometries and attributes, raster pixels, table rows, table names, column schemas, feature counts, extents, source URLs and filenames, contacts, and sampled row values. This issue is fixed in version 1.2.3.

Summary

Multiple GeoLens read/link endpoints authorized only the resource named in the request URL (a map, a VRT, a source dataset, an AI request) and failed to re-authorize a second, caller-influenced dataset that the request reached through a relationship, layer reference, mosaic source, or request body. This "authorize the URL resource, read a different dataset un-re-authorized" pattern let callers read data from datasets they have no access to.

The most severe instances require no authentication at all (anonymous, network-only). Others require only the default editor role that any self-service signup / upload user receives.

All issues are fixed in 1.2.3. There is no complete configuration workaround — upgrading is the only full remediation.

Impact

Depending on the endpoint, an attacker can read, for datasets they cannot otherwise access:

• the dataset's vector tile data (actual feature geometries/attributes),
• the dataset's raster pixels,
• backing-table rows,
• and metadata — table name, column schema, feature count, extent, source URL/filename, contacts, and sampled row values.

Affected versions

All versions prior to 1.2.3 (includes the published 1.0.0, 1.2.0, and 1.2.2 releases and their PyPI/npm/GHCR artifacts). Fixed in 1.2.3.

Findings

1. Anonymous metadata + private vector-tile disclosure via public maps (PR #235) GET /maps/{id} and GET /maps/{id}/style.json authorized the map but not each layer's backing dataset. A public map that references a private dataset leaked that dataset's table name, column schema, feature count, extent, and sampled values to anonymous callers. style.json additionally returned a vector-tile URL carrying an HMAC signature bound to neither user nor map, which the tile endpoint accepts for non-public datasets with no user check — so the signature is replayable to read the private dataset's actual vector tiles. (Anonymous · High)

2. Anonymous private-row disclosure via dataset relationships (PR #234) The dataset FK-relationship APIs authorized only the source dataset from the URL, never the relationship target. A public dataset with a relationship to a private dataset let an anonymous caller enumerate the relationship (obtaining the private target's id/title and the relationship id) and then call the related-record endpoint to read rows from the private target's backing table. (Anonymous · High)

3. Anonymous metadata disclosure via OGC externalId lookup (PR #236) GET /collections/datasets/items?externalId=<uuid> resolved the dataset by id and returned the full OGC catalog record (title, summary, bbox, keywords, contacts, distributions, source org) with no visibility check — the user was never threaded into the lookup. An anonymous caller could read any private, restricted, or unpublished dataset's metadata by UUID. (Anonymous · High)

4. Cross-tenant raster pixel disclosure via VRT mosaics — SEC-C (PR #237) An authenticated user with the default editor upload permission could mosaic another user's private raster into a VRT they own, then read the victim's pixels back through raster tile / quicklook / COG endpoints that authorize only the attacker-owned VRT. VRT member pixels are compiled into one served asset and cannot be filtered at read time, so the fix authorizes every source dataset at write/link time. (Authenticated editor · High)

5. Cross-tenant metadata/sample-data disclosure via AI metadata endpoints — SEC-D (PR #238) The POST /ai/metadata/{summary,keywords,lineage,quality-statement} endpoints were gated only by the use_ai_chat permission (held by the default editor role). The attacker-controlled dataset_id in the request body flowed into the LLM prompt context with no visibility filter, rendering any dataset's title, summary, source URL, filename, column schema, and sample values into the response.…

Summary

Multiple GeoLens read/link endpoints authorized only the resource named in the request URL (a map, a VRT, a source dataset, an AI request) and failed to re-authorize a second, caller-influenced dataset that the request reached through a relationship, layer reference, mosaic source, or request body. This "authorize the URL resource, read a different dataset un-re-authorized" pattern let callers read data from datasets they have no access to.

The most severe instances require no authentication at all (anonymous, network-only). Others require only the default editor role that any self-service signup / upload user receives.

All issues are fixed in 1.2.3. There is no complete configuration workaround — upgrading is the only full remediation.

Impact

Depending on the endpoint, an attacker can read, for datasets they cannot otherwise access:

• the dataset's vector tile data (actual feature geometries/attributes),
• the dataset's raster pixels,
• backing-table rows,
• and metadata — table name, column schema, feature count, extent, source URL/filename, contacts, and sampled row values.

Affected versions

All versions prior to 1.2.3 (includes the published 1.0.0, 1.2.0, and 1.2.2 releases and their PyPI/npm/GHCR artifacts). Fixed in 1.2.3.

Findings

1. Anonymous metadata + private vector-tile disclosure via public maps (PR #235) GET /maps/{id} and GET /maps/{id}/style.json authorized the map but not each layer's backing dataset. A public map that references a private dataset leaked that dataset's table name, column schema, feature count, extent, and sampled values to anonymous callers. style.json additionally returned a vector-tile URL carrying an HMAC signature bound to neither user nor map, which the tile endpoint accepts for non-public datasets with no user check — so the signature is replayable to read the private dataset's actual vector tiles. (Anonymous · High)

2. Anonymous private-row disclosure via dataset relationships (PR #234) The dataset FK-relationship APIs authorized only the source dataset from the URL, never the relationship target. A public dataset with a relationship to a private dataset let an anonymous caller enumerate the relationship (obtaining the private target's id/title and the relationship id) and then call the related-record endpoint to read rows from the private target's backing table. (Anonymous · High)

3. Anonymous metadata disclosure via OGC externalId lookup (PR #236) GET /collections/datasets/items?externalId=<uuid> resolved the dataset by id and returned the full OGC catalog record (title, summary, bbox, keywords, contacts, distributions, source org) with no visibility check — the user was never threaded into the lookup. An anonymous caller could read any private, restricted, or unpublished dataset's metadata by UUID. (Anonymous · High)

4. Cross-tenant raster pixel disclosure via VRT mosaics — SEC-C (PR #237) An authenticated user with the default editor upload permission could mosaic another user's private raster into a VRT they own, then read the victim's pixels back through raster tile / quicklook / COG endpoints that authorize only the attacker-owned VRT. VRT member pixels are compiled into one served asset and cannot be filtered at read time, so the fix authorizes every source dataset at write/link time. (Authenticated editor · High)

5. Cross-tenant metadata/sample-data disclosure via AI metadata endpoints — SEC-D (PR #238) The POST /ai/metadata/{summary,keywords,lineage,quality-statement} endpoints were gated only by the use_ai_chat permission (held by the default editor role). The attacker-controlled dataset_id in the request body flowed into the LLM prompt context with no visibility filter, rendering any dataset's title, summary, source URL, filename, column schema, and sample values into the response.…

Summary

Multiple GeoLens read/link endpoints authorized only the resource named in the request URL (a map, a VRT, a source dataset, an AI request) and failed to re-authorize a second, caller-influenced dataset that the request reached through a relationship, layer reference, mosaic source, or request body. This "authorize the URL resource, read a different dataset un-re-authorized" pattern let callers read data from datasets they have no access to.

The most severe instances require no authentication at all (anonymous, network-only). Others require only the default editor role that any self-service signup / upload user receives.

All issues are fixed in 1.2.3. There is no complete configuration workaround — upgrading is the only full remediation.

Impact

Depending on the endpoint, an attacker can read, for datasets they cannot otherwise access:

• the dataset's vector tile data (actual feature geometries/attributes),
• the dataset's raster pixels,
• backing-table rows,
• and metadata — table name, column schema, feature count, extent, source URL/filename, contacts, and sampled row values.

Affected versions

All versions prior to 1.2.3 (includes the published 1.0.0, 1.2.0, and 1.2.2 releases and their PyPI/npm/GHCR artifacts). Fixed in 1.2.3.

Findings

1. Anonymous metadata + private vector-tile disclosure via public maps (PR #235) GET /maps/{id} and GET /maps/{id}/style.json authorized the map but not each layer's backing dataset. A public map that references a private dataset leaked that dataset's table name, column schema, feature count, extent, and sampled values to anonymous callers. style.json additionally returned a vector-tile URL carrying an HMAC signature bound to neither user nor map, which the tile endpoint accepts for non-public datasets with no user check — so the signature is replayable to read the private dataset's actual vector tiles. (Anonymous · High)

2. Anonymous private-row disclosure via dataset relationships (PR #234) The dataset FK-relationship APIs authorized only the source dataset from the URL, never the relationship target. A public dataset with a relationship to a private dataset let an anonymous caller enumerate the relationship (obtaining the private target's id/title and the relationship id) and then call the related-record endpoint to read rows from the private target's backing table. (Anonymous · High)

3. Anonymous metadata disclosure via OGC externalId lookup (PR #236) GET /collections/datasets/items?externalId=<uuid> resolved the dataset by id and returned the full OGC catalog record (title, summary, bbox, keywords, contacts, distributions, source org) with no visibility check — the user was never threaded into the lookup. An anonymous caller could read any private, restricted, or unpublished dataset's metadata by UUID. (Anonymous · High)

4. Cross-tenant raster pixel disclosure via VRT mosaics — SEC-C (PR #237) An authenticated user with the default editor upload permission could mosaic another user's private raster into a VRT they own, then read the victim's pixels back through raster tile / quicklook / COG endpoints that authorize only the attacker-owned VRT. VRT member pixels are compiled into one served asset and cannot be filtered at read time, so the fix authorizes every source dataset at write/link time. (Authenticated editor · High)

5. Cross-tenant metadata/sample-data disclosure via AI metadata endpoints — SEC-D (PR #238) The POST /ai/metadata/{summary,keywords,lineage,quality-statement} endpoints were gated only by the use_ai_chat permission (held by the default editor role). The attacker-controlled dataset_id in the request body flowed into the LLM prompt context with no visibility filter, rendering any dataset's title, summary, source URL, filename, column schema, and sample values into the response.…