🔥 CyberForge CVE of the Day #105
🚨 CVE-2026-129171 — NginxPulse: when statistics selectors become SQL
🎯 The quick hit:
NginxPulse, the application that turns NGINX logs into analytics, has a critical SQL-injection flaw in its statistics API. The CNA’s range is 1.5.0 up to, but excluding, 1.6.26. Request-controlled id and locationType values can cross into SQL structure. In the researcher’s account, no credentials are needed by default; if system access keys are configured, the caller needs a key.
Establish the serving NginxPulse build, route access history and actual PostgreSQL identity. An NGINX banner alone does not establish scope. The researcher’s official-image setup used a database superuser, materially increasing potential impact.
Prioritise exposed affected services, preserve available request/application/database records and move to a supported fixed artifact. A supported access boundary can reduce reachability while repair is pending, but access-key enforcement does not remove the SQL defect. Embedded PostgreSQL upgrade planning must account for role migration and backup; external database roles need their own review.
Public technical reproduction exists. This briefing did not execute it. The captured KEV catalog has no exact entry, and the latest returned EPSS probability is 0.414% for 15th November. Those are dated signals, not a clean-system result or permission to ignore an applicable critical flaw.
🔑 Key details:
• Identity: CVE-2026-129171; PUBLISHED; CNA VulnCheck; publisher likaia; product NginxPulse. Credit: Wu Yulun.
• Surface: /api/stats/:type; id and locationType parameters. Weakness: CWE-89, SQL Injection.
• Published affected range: >=1.5.0,<1.6.26. The original researcher’s stated range ends at 1.6.24; the CNA’s range includes 1.6.25. Keep that distinction visible rather than assuming the narrower report excludes the later version.
• Repair: v1.6.26. Its source tag resolves to corrective commit f9416bea3be892a394cd24a6d4a2afbd8397e926. Source identity is not proof that a downloaded binary or running image contains the fix.
• CNA CVSS v4.0: 9.3 CRITICAL — CVSS:4.0/AV:N/AC:L/AT:N/PR:N/UI:N/VC:H/VI:H/VA:H/SC:N/SI:N/SA:N.
• CNA CVSS v3.1: 9.8 CRITICAL — CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H.
• NVD status: Received. Captured metrics are attributed to VulnCheck, with v3.1 Primary and v4.0 Secondary entries. No separate NVD-authored score was present. Its v4 string expands undefined optional metrics; that does not make it a new independent assessment.
• KEV: exact non-match in the complete captured 1,781-row catalog, version 2026.11.13, checked 16th November.
• FIRST EPSS: probability 0.004140000 (0.414%); percentile 0.362610000 (36.261%); score date 2026-11-15; retrieved 16th November. Percentile is a relative rank, not the probability.
🧬 What actually went wrong?
The site selector should choose a configured site and the location selector a known reporting mode. Instead, they influence a quoted table identifier or column expression. SQL placeholders protecting other values do not validate those separately constructed structural positions.
The researcher describes error-based disclosure through the site-ID path and a direct output path through location statistics. The CNA describes database-content access and server-file reads/writes through powerful PostgreSQL functions. The issue also discusses further execution consequences under its deployment conditions. Keep attribution and privilege boundaries attached: this review did not reproduce those effects, establish a reliable universal exploit or prove host-root access.
The corrective code does several useful things. It validates website IDs against the configuration’s known site registry, including runtime collision-suffixed IDs. It constrains locationType to domestic, city or global and rejects unsupported values. It replaces detailed query-error responses with a generic HTTP error while retaining diagnostic context in server logs. Describing the fix as only hiding an error message would miss the input-boundary repairs.
There is also a database-privilege layer. The patch inspects direct and reachable role power, warns about excessive privileges or inspection failure and separates administrator/application roles for embedded PostgreSQL. A warning is configuration evidence, not evidence of an attacker. Startup continues for compatibility; external DSN accounts are not automatically demoted. The SQL fixes do not depend on that demotion.
The important consequence boundary is the PostgreSQL server. Its filesystem, operating-system account and mounts may be inside a container or on an external database host. A database superuser does not automatically identify an OS root identity or demonstrate container escape. Broader host/process claims need their own evidence.
📦 Which deployments belong in scope?
Inventory NginxPulse’s active release/artifact, every replica and its deployment interval. A fixed checkout, unused image or updated tag does not establish the serving process. Reconcile release/digest provenance with the active instance.
Use the CNA range, including 1.6.25 unless vendor evidence establishes a particular repaired artifact. OSV’s generated commit range and the unreviewed GHSA mirror are derivative context, not independent range validation.
Establish historical direct/proxied routes, edge controls, system.accessKeys and potential key holders. Protect keys and credential-bearing DSNs. Current policy cannot reconstruct earlier exposure; a configured key does not make the code unaffected.
Finally split embedded PostgreSQL from external DB_DSN deployments. The database identity, inherited/SET ROLE privileges, server location, mounts and file permissions determine how far an injected query could reach. These are local inventory questions, all unknown in this research package. A powerful role warrants review; it proves neither exploitation nor that every adjacent service was affected.
🕰️ Exposure window and timeline:
The public issue was created on 30th October at 09:29:06. The corrective commit was authored on 4th November at 11:06:39 and committed at 11:06:48; v1.6.26 was released at 11:10:00. The maintainer’s acknowledgement followed that day. The CVE was published on 13th November at 17:21:48, and NVD received it later that day.
That is a disclosure/repair timeline, not a local exposure window. An installation may have served affected code before the public report. Reconstruct deployment, route and key-policy changes; end the interval only when the fixed serving state or relevant access boundary is supported by evidence. Explicitly record any portion outside available log retention.
EPSS is dated 15th November; retrieval/cutoff is 16th November; the KEV catalog release is 13th November. A later refresh needs a new cutoff and review of dependent claims.
🗺️ MITRE ATT&CK:
T1190 — Exploit Public-Facing Application, Initial Access — is a conditional CyberForge mapping for an Internet-facing affected NginxPulse statistics service that is exploited through this SQL injection. The behaviour basis is the original research/issue; the technique definition is MITRE’s current page, checked 16th November. Local Internet-facing exposure is still an evidence question.
Pivot from historical build/access to requests, app errors, DB sessions and corroborated effects. A CVE match or query punctuation does not establish initial access; internal-only services need their own context.
No execution, persistence, lateral movement, exfiltration or actor/campaign mapping is claimed here. A possible consequence discussed by a researcher is not proof that a later intrusion stage happened. CWE-89 supplies the weakness classification, not a complete ATT&CK chain. Keep the reference and the behaviour evidence separate.
🧭 SOC investigation trail:
Use these six proposed checks as a handover. Keep exposure, suspicion, corroborated unauthorised effects and unavailable telemetry separate. No task is marked complete here.
C01 — Scope the serving application
Action: inventory each active build, route and key policy. Location: deployment history, app/proxy configuration and serving process/container. Signal: >=1.5.0,<1.6.26 during a reachable interval. Correlate: route, restart and key-policy changes. Limits: an NGINX banner, unused image or access key does not establish a repaired app. Escalate: affected reachability, unknown build or undocumented access needs an owner and deadline.
C02 — Preserve the affected interval
Action: arrange bounded collection before rotation, restarts or migration, subject to urgent containment. Location: available proxy/app/PG logs, configuration, deployment exports and approved backups. Signal: records with original timestamps and provenance. Correlate: clocks, instances, database host and retention. Limits: query strings, SQL or bodies may be absent; protect secrets/personal data. Escalate: imminent loss, failed collection or uncovered time; retain gaps explicitly.
C03 — Hunt for request-to-SQL anomalies
Action: examine retained statistics-route id/locationType inputs without sending payloads. Location: proxy records, app query-error context and PG errors if collected. Signal: unusual selector structure with related conversion/query failures. Correlate: request time/ID, route, site, instance and DB session where available. Limits: malformed clients, stale sites, setup changes and authorised tests can explain errors; HTTP500 proves no outcome. Escalate: a coherent sequence with corroborating DB behaviour.
C04 — Bound privilege and consequence
Action: have the DBA establish actual app login/effective roles and inherited/SET ROLE power. Location: DSN identity, PG role/audit records, DB host/container and mounts. Signal: powerful access plus independently supported unusual statements or data/file/process effects. Correlate: original requests, DB session, filesystem permissions and lineage. Limits: a superuser/warning is configuration evidence; DB-server files do not establish app-host/root compromise. Escalate: corroborated unauthorised effects to incident response.
C05 — Repair under the correct database plan
Action: select a supported fixed artifact, preserve rollback material and reduce reachability while pending. Location: release/provenance, deployment records and embedded/external guidance. Signal: approved 1.6.26-or-later repair plan; embedded pgdata backup and role-migration ownership. Correlate: change logs, administrator/app separation and restart results. Limits: external roles are untouched automatically; SQL repair does not require demotion. Escalate: migration errors, unsupported rollback or suspected unauthorised changes.
C06 — Verify and hand over
Action: confirm every serving instance, normal analytics, intended access and remaining evidence questions. Location: active artifact/process mapping, restarts, configuration and case records. Signal: provenance and running state agree; applicable role/migration review is recorded. Correlate: replicas, failover, stale jobs and pre-change interval. Limits: stored images or absence of errors prove neither a running repair nor historical cleanliness. Escalate: stale copies, failed health, unexplained changes or gaps preventing a defensible disposition.
Record the owner, asset, UTC interval, evidence references and next decision. A suspicious request, a correlated database event and a corroborated unauthorised effect support different conclusions. No validated malicious IOC set or local incident is established. Missing evidence remains a visibility gap, even after an upgrade.
🔎 Indicators and observable behaviour:
The useful leads are contextual: the affected statistics route, structurally unusual id/locationType values, related query/conversion errors and independently supported database/file/process effects. None is a standalone compromise signature. Normal analytics uses the same route; stale site IDs, malformed clients, setup changes and authorised testing can explain failures.
The fixed startup events security_event=postgres_role_privileges and security_event=postgres_role_check_failed concern privilege inspection. The first can flag a powerful role; the second means inspection did not complete. They do not show exploitation, and absence of a warning cannot replace a role review. Generic HTTP500 output also cannot fingerprint the fixed release reliably.
Query strings, SQL, bodies and role changes may not have been logged. Record missing fields and joins explicitly. No findings ≠ proof of a clean system.
🧰 Optional CVE checks:
These are source-checked, read-only examples, not commands run by CyberForge. Use the authorised owner/DBA, choose the intended asset/project/connection and time-bound the evidence you retain. Access errors and missing output remain unknowns.
Q01 — Created-container image inventory:
```bash
docker compose images
```
Docker documents this as listing images used by created containers in the selected Compose project. Run only in that project’s established management context. Reconcile the returned service/image identifiers with release/digest and deployment records. It does not cover other projects or standalone binaries, prove a listed container is serving, attest a mutable tag or verify every replica. It is an inventory lead, not a CVE test.
Q02 — Direct flags for the actual application session/login identities:
```sql
SELECT current_user AS effective_role,
session_user AS login_role,
r.rolname,
r.rolsuper, r.rolcreatedb, r.rolcreaterole,
r.rolreplication, r.rolbypassrls
FROM pg_catalog.pg_roles AS r
WHERE r.rolname IN (current_user, session_user);
```
Use a DBA-approved PostgreSQL connection representing the actual app identity, with approved credential handling and readable catalog metadata. An unrelated administrator session reports that administrator. Expect one or two role rows; powerful flags warrant design review. This query does not enumerate memberships, inherited/SET ROLE power, server-file/program roles, grants, OS permissions or historical role state. False direct flags therefore do not establish least privilege. The fixed app’s inspection is broader and recursive.
Both examples were checked against official documentation; neither was executed or validated against representative data. No injection, file primitive, role mutation, migration or disposable-container security-test script was run. Keep operational output containing topology or identities in the restricted case record.
🩹 Remediation and verification:
Upgrade to 1.6.26 or a later supported artifact containing the correction, with traceable provenance. Restrict untrusted statistics access while pending and verify the actual control boundary, including forwarded/direct paths. Do not treat an enabled access key as the patch. Preserve available evidence according to urgency and the incident procedure before disruptive changes erase it.
Embedded PostgreSQL needs explicit planning. The fixed deployment guide separates an administrator, default nginxpulse_admin, from the application account. Existing-cluster migration renames the old bootstrap superuser, recreates an ordinary application login and transfers object ownership. Back up pgdata before upgrading. The documentation warns against directly rolling back to an old image after migration: the old entrypoint expects database-creation power on the application identity. Those are vendor design statements, not a migration tested here.
External PostgreSQL / DB_DSN behaves differently. The upgrade leaves external roles and memberships unchanged. The SQL fixes are independent of role demotion, and privilege hardening is separate DBA work, not an upgrade prerequisite. Identify the actual login/effective roles and direct/indirect power; a bootstrap superuser requires a supported separate ordinary-app-role design rather than an assumed simple demotion.
Acceptance needs every serving artifact, completed restarts, intended access, normal analytics and applicable backup/migration/role records. A stored image, generic error or absent alert is insufficient.
Keep historical incident closure separate. A patch does not restore altered data or settle credential response/recovery. Hand over findings, interval, retained/lost evidence, gaps and owner. No remediation or recovery was performed here.
🧾 Evidence separation:
CNA statements supply identity, range, scores and impact assessment. Wu Yulun’s report supplies attributed reproduction and access conditions. The project commit/tag/files supply passive repair and deployment evidence. NVD repeats source-labelled CNA metrics; OSV and the unreviewed GHSA mirror are derivative context. KEV and EPSS are dated catalog/model snapshots.
CyberForge supplies the conditional ATT&CK mapping, prioritisation and proposed evidence trail. No local asset, database, exploit, source build, detector, migration or incident was tested. Source-reviewed maintainer tests are not locally passed tests. This is a reviewed local publishing draft; platform/composer acceptance and posting are separate activities.
🔥 CyberForge verdict:
Give reachable affected NginxPulse services urgent ownership, especially where the app’s database connection has powerful privileges. The basis is a critical request-to-SQL failure, default no-key access in the research account, public technical detail and a published correction. A low forecast and a catalog non-match do not undo that consequence if the deployment is applicable.
The practical route is clear: identify the serving app, establish access history, preserve useful records, repair under the right database-mode plan and verify the running result. Review database power and investigate supported anomalies with the correct server boundary. Keep the vulnerability repair and historical incident disposition independently defensible.
🔗 Sources and further reading:
• VulnCheck CNA: https://www.vulncheck.com/advisories/nginxpulse-1.5.0-before-1.6.26-sql-injection-via-api-stats-type
• Original research, Wu Yulun: https://vul.yulun.ac.cn/20261030/nginxpulse/
• Maintainer issue: https://github.com/likaia/nginxpulse/issues/103
• Corrective commit: https://github.com/likaia/nginxpulse/commit/f9416bea3be892a394cd24a6d4a2afbd8397e926
• Fixed release: https://github.com/likaia/nginxpulse/releases/tag/v1.6.26
• Deployment/migration guide: https://github.com/likaia/nginxpulse/blob/v1.6.26/docs/fumadocs/content/docs/Deployment-EN.md
• NVD: https://nvd.nist.gov/vuln/detail/CVE-2026-129171
• KEV: https://www.cisa.gov/known-exploited-vulnerabilities-catalog
• EPSS: https://api.first.org/data/v1/epss?cve=CVE-2026-129171
• ATT&CK: https://attack.mitre.org/techniques/T1190/
• Command references: https://docs.docker.com/reference/cli/docker/compose/images/