🔥 CyberForge CVE of the Day #106
🚨 CVE-2026-104840 — OpenCTI: repeated login attempts and conditional account access
🎯 The quick hit:
OpenCTI’s primary login uses the GraphQL token mutation. Before 7.260428.0, it lacked the endpoint-specific restriction described in the advisory, while the global Express limiter defaulted to 10,000 requests per one-second window. That broad ceiling is not a measured password-guessing rate or a useful promise about account-login protection.
An unauthenticated caller can repeatedly try passwords against an account. The consequential step is conditional: a usable credential must be recovered and account authentication/use must succeed under the actual controls. This is not passwordless login. The permitted data and operations depend on the recovered account; administrator rights, all CTI data and host execution are not automatic outcomes.
Inventory the actual serving backend and primary-login reachability, preserve available authentication context and apply a supported fixed artifact. Review proxy identity, exemptions and both limiter layers. Investigate suspicious account use separately from the software change.
Two source boundaries matter today. The project labels the advisory Moderate, while the CNA supplies a 9.2 Critical CVSS v4.0 assessment. FIRST returned no EPSS row: unavailable, not zero. The correction was released in April, before the July project advisory date and November CVE publication. Check what is already serving before planning a change.
🔑 Key details:
• Identity: CVE-2026-104840; PUBLISHED; CNA GitHub_M; publisher OpenCTI-Platform; product opencti / OpenCTI. Advisory GHSA-74jw-q47r-q5pv. Reporter credit: jeroengui.
• Weakness: CWE-307 — Improper Restriction of Excessive Authentication Attempts. Surface: primary-login token GraphQL mutation.
• Affected: literal CNA row “< 7.260428.0”; no lower bound or defaultStatus supplied. Project patched scope: >=7.260428.0. Do not invent a starting release or compare this calendar-like version lexically.
• CNA CVSS v4.0: 9.2 CRITICAL — CVSS:4.0/AV:N/AC:H/AT:P/PR:N/UI:N/VC:H/VI:H/VA:H/SC:N/SI:N/SA:N.
• AC:H and AT:P are explicit. Do not replace them with low complexity or no attack conditions. PR:N/UI:N describe starting access, not guaranteed credential success.
• Project advisory: Moderate, with no separate numeric vector on the captured page. Preserve the difference; do not average or silently replace it.
• NVD: Received, with a Secondary metric attributed to security-advisories.com; no independent NVD-authored score captured. Its expanded undefined optional v4 fields do not create another assessment. No CVSS v3.1 score was supplied.
• Repair: 7.260428.0 source tag resolves to 8a99a8f461a2faf9fc9c12a8e92cd32e2935a40d. Both linked corrective commits are ancestors, and the fixed-tag schema contains the login directive. Source ancestry is not binary or running-image attestation.
• KEV: exact non-match in the complete captured 1,781-row catalog, version 2026.11.13, checked 17th November. Not a local safety result.
• FIRST EPSS: successful exact request returned total:0, data:[]. Probability, percentile and score date are unavailable; no 0% or low forecast is asserted.
🧬 What actually went wrong?
Checking one password correctly does not protect an account from excessive repeated checks. The flaw is the missing or ineffective restriction on attempts at primary login. The documented consequence is access if guessing recovers usable credentials, bounded by the account and any applicable authentication controls.
The repair has two layers. Commit 8599366319532dbdc1cfd553c55a518ad146c3df moves the global Express limiter to the first middleware registration, before sessions and GraphQL. The shared builder keeps a default of 10,000 requests/1,000ms and adds an exact-string req.ip skip list. Its handler returns HTTP429 and emits a debug-level over-quota message.
Commit d7dfab4c264e36addf215b413924cb77f0827070 adds (limit: 1, duration: 1) to token. The fixed schema confirms it. OpenCTI wires the directive’s onLimit hook to throw a FunctionalError. A GraphQL rejection therefore needs its actual error/outcome interpretation; HTTP429 is not the only possible evidence, and HTTP200 does not prove every GraphQL operation succeeded.
There is a useful source-derived limit on what to promise. OpenCTI’s reviewed wiring only overrides onLimit. The pinned graphql-rate-limit-directive 2.0.6 source defaults to a schema-coordinate key and RateLimiterMemory. That supports a field-level, in-process interpretation, not a demonstrated per-account lockout or shared counter across replicas. This was passive source analysis, not a measured capacity problem or a new bypass finding.
📦 Which deployments belong in scope?
Identify the backend actually serving primary login, its active release/artifact and every replica. A stored fixed image, frontend label, connector/worker component or updated checkout alone does not establish the serving authentication code. Include restart and deployment history for the relevant interval.
Use the published range before 7.260428.0, with vendor/fork evidence for exceptions. The generated OSV Git range and version list are derivative context, not an independent experiment covering every historical release. The fixed tag includes both corrections; a particular distributed binary/image still needs provenance and runtime mapping.
Establish direct and proxied reachability, private/VPN boundaries, native credential policy and the actual SSO/MFA path. An SSO-facing screen does not prove native token login is unreachable. This review proves neither universal MFA bypass nor universal protection. Current configuration cannot reconstruct previous access conditions by itself.
Then review global settings, IP exemptions, trusted forwarding and replica/counter scope. Shared NAT or a proxy can collapse several legitimate users into one apparent client. The global skip list affects the reviewed Express layer; it does not automatically configure the separate login-field directive. Actual configuration and workload are unknown here.
🕰️ Exposure window and timeline:
Issue15666 opened on 23rd April. The middleware/configuration correction was committed 27th April at 11:19:28. Issue15735 and the login correction followed on 28th April; the latter committed at 06:47:07. The 7.260428.0 release was published on 28th April at 15:03:22.
The project advisory displays 28th July 2026. The CVE record says reserved 2nd October and published 16th November at 21:00:01; NVD received it later that evening. Keep those source dates separate. The project publication preceding the displayed reservation date is retained without an invented disclosure explanation.
A local exposure interval starts with supported affected deployment and relevant reachability, not CVE publication. End it only when fixed serving state or the changed boundary is established. Preserve deployment/provider/route history and identify any period outside log retention. This 17th November source snapshot does not establish when a client installed, exposed or repaired the service.
🗺️ MITRE ATT&CK:
T1110.001 — Password Guessing, Credential Access — is a conditional CyberForge mapping of the repeated attempts described by the CNA/project. Pivot to real authentication failures by account/time, then trustworthy client context, instance and rate pattern. A GraphQL request count alone does not establish a failed password check or malicious guessing.
T1078 — Valid Accounts, Initial Access when applicable — is conditional only after recovered credentials are successfully used for unauthorised access. Pivot to successful authentication/session evidence and subsequent account actions under actual permissions. Password recovery, a normal user’s success or temporal proximity to failures does not by itself establish malicious use.
MITRE’s current technique pages were checked 17th November; the case behaviour comes separately from the advisory/CNA. No local ATT&CK finding, actor, campaign, persistence, exfiltration method or host-execution chain is asserted. CWE-307 classifies the weakness, not an entire intrusion.
🧭 SOC investigation trail:
Six proposed checks follow. Keep applicability, exposure, suspicion, corroborated unauthorised effects and unavailable telemetry separate. These are evidence questions, not completed operational tasks.
C01 — Scope the serving build and login path
Action: inventory each active release, replica, primary-login route and identity controls. Location: deployment history, app/proxy policy and runtime records. Signal: before 7.260428.0 with relevant reachability. Correlate: upgrades, restarts, direct/proxied paths and provider changes. Limits: a fixed stored image or SSO screen proves neither serving repair nor native-path history. Escalate: affected exposure, unknown build or undocumented access needs an owner.
C02 — Preserve context without credentials
Action: collect bounded available evidence before rotation/restart, subject to urgency. Location: proxy/app/audit/identity records and configuration/deployment exports. Signal: original times, outcomes, account/client context and provenance. Correlate: clocks, transports, replicas and retention. Limits: ./logs and retained-file defaults do not guarantee a complete login history; redact passwords/tokens. Escalate: imminent loss, missing outcome fields or uncovered intervals; retain the gap.
C03 — Hunt for repeated failed logins
Action: review confirmed primary-login failures by target account/time without sending guesses. Location: verified authentication/audit fields and related proxy/app records. Signal: a coherent suspicious failure pattern. Correlate: account, trusted client attribution, route, instance and existing alerts. Limits: forgotten passwords, stale automation, NAT and sanctioned tests can explain activity; GraphQL traffic or status alone proves no credential failure. Escalate: a supported suspicious sequence after benign/coverage review.
C04 — Correlate success and account actions
Action: examine nearby successful authentication and later activity. Location: available identity/session, audit and CTI records with confirmed schema. Signal: unexplained successful account use or a sensitive operation tied to the interval. Correlate: account, sessions, trusted client context, normal behaviour and permissions. Limits: success after failures can be legitimate; IP/proximity alone proves no attacker or data loss. Escalate: corroborated unauthorised use/effects to incident response.
C05 — Repair both limiter layers
Action: select supported fixed code and review global/login-field controls. Location: release/provenance, two corrections, active policy and proxy topology. Signal: correction included in 7.260428.0-or-later serving plan with intended identity/exemptions. Correlate: first HTTP middleware, token directive, global settings, skip list and replicas. Limits: lowering a ceiling or enabling logging is not completed repair; shared proxy identities can affect users. Escalate: stale replicas, unexplained exemptions or auth availability problems.
C06 — Verify and hand over separately
Action: confirm each serving artifact and intended normal auth behaviour; assign historical gaps. Location: runtime/deployment, policy and case records. Signal: provenance/running state agree and normal login remains usable. Correlate: replicas, restarts, direct paths and pre-change interval. Limits: reviewed defaults are field-scoped/in-memory, not proven per-account or shared-cluster protection; no live test ran here. Escalate: ineffective controls, unexplained success or gaps preventing a defensible disposition.
Handover the owner, asset/account, UTC interval, evidence references, benign explanations, coverage limits and next decision. A suspicious failure pattern, successful authentication and unauthorised CTI operation support different conclusions. Missing evidence remains an explicit limitation after an upgrade.
🔎 Indicators and observable behaviour:
No validated malicious IP, domain, hash, actor or campaign IOC set is supplied by the retained primary sources. Contextual leads are repeated confirmed primary-login failures, unexplained success and account operations that can be tied to the interval. Forgotten passwords, stale automation, shared egress, normal analyst work and sanctioned testing need consideration.
The global debug message begins [RATE-LIMIT] over quota for; its HTTP handler returns 429. It is general traffic evidence, not a login-specific signature. Default app logging is info, so a missing debug line does not establish that limiting never occurred. The GraphQL directive’s Too many requests FunctionalError follows a separate path. Verify the actual outcome schema and transport.
Do not capture passwords, tokens or raw login bodies merely to make a detector easier. Default redacted-input settings are source evidence, not proof every custom proxy or log transport redacts correctly. No findings ≠ proof of a clean system.
🧰 Optional CVE checks:
Both examples are documentation-checked and unexecuted. Use the authorised management context, known backend container and justified interval. They do not test passwords or limiter effectiveness; errors and absent records remain unknowns.
Q01 — Created-container image inventory:
```bash
docker compose images
```
Docker lists image information for containers created in the selected Compose project. Reconcile service/image identifiers with release/digest, actual backend, replicas and restart records. This excludes other projects and standalone binaries; a listed image or mutable tag does not establish loaded code, current serving state or that every instance is fixed. Output may disclose topology.
Q02 — One known backend’s bounded console records:
```bash
docker logs --since '2026-11-16T00:00:00Z' --until '2026-11-17T08:55:07Z' --timestamps --tail 200 OPENCTI_CONTAINER
```
Replace OPENCTI_CONTAINER with the known backend name/ID and the example times with the supported investigation bounds. These dates are not an asserted client exposure interval. Expect available stdout/stderr records or an error, subject to Docker permissions and logging driver. The 200-line tail cap, rotation and source configuration can omit evidence. File-only audit records are not collected by this command. Default info may omit the global limiter’s debug diagnostic; an empty result does not establish historical cleanliness. Protect identities, tokens and sensitive CTI context.
Neither command was run against representative data or a deployment. No login request, password list, guessing loop, rate benchmark, account mutation, token operation or detector was executed. A useful SIEM query needs the real schema and coverage; this article does not invent authentication event fields.
🩹 Remediation and verification:
Upgrade affected services to 7.260428.0 or later supported software containing the correction, with traceable provenance and the owner’s supported procedure. Some systems may already have the April repair; verify what is serving. Restrict untrusted primary-login reachability while pending through supported controls and preserve evidence according to urgency/incident procedure.
Review APP__RATE_PROTECTION__MAX_REQUESTS, APP__RATE_PROTECTION__TIME_WINDOW and APP__RATE_PROTECTION__IP_SKIP_LIST against workload and active configuration. The source skip list matches exact IP strings; do not invent CIDR semantics. Review the actual proxy trust/hop chain and direct paths before interpreting req.ip. Do not broaden trust or weaken authentication as a shortcut; shared identity can also affect legitimate-user availability.
Confirm both repair layers, all serving artifacts and required restarts. Review login-field key/store/process scope and normal authentication usability through approved owner checks. Schema limit 1/duration 1 is not proof of per-account or cluster-wide state. Avoid generating production password guesses to demonstrate a fix; behavioural testing belongs in an approved environment.
Keep historical incident closure separate. If unauthorised account/data effects are corroborated, the incident owner should determine credential/session response, permission/data integrity and recovery scope. An upgraded container does not settle those questions. No mitigation, account reset, session revocation or recovery was performed for this briefing.
🧾 Evidence separation:
The CNA supplies exact identity, range, CWE and 9.2 vector. The project supplies its Moderate label, reporter credit, conditional impact statement and displayed publication date. Repository diffs, fixed-tag content and ancestry supply passive repair evidence. Pinned dependency source supports an implementation inference, not a deployment measurement. NVD repeats source-labelled scoring; OSV is generated context.
KEV is a dated exact catalog non-match; EPSS has no row. The global GHSA API returned 404 while project HTML remained readable, which does not prove withdrawal. CyberForge supplies proposed SOC checks, conditional ATT&CK mappings and prioritisation. No local asset, login, credential, rate test, detector, incident or publication was assessed operationally.
🔥 CyberForge verdict:
Give reachable affected OpenCTI primary-login services prompt repair ownership, especially where native credentials and sensitive CTI permissions are in use. The basis is the documented repeated-attempt weakness, conditional account consequences and a published correction. Actual urgency still needs build, exposure, identity controls and business context.
Identify the serving backend, preserve useful outcomes, repair and verify both limiter layers under the real proxy/replica model. Investigate supported account anomalies with privacy and attribution intact. A KEV non-match or unavailable score establishes no safety; a critical score establishes no local incident.
🔗 Sources and further reading:
• Project advisory: https://github.com/OpenCTI-Platform/opencti/security/advisories/GHSA-74jw-q47r-q5pv
• CNA record: https://cveawg.mitre.org/api/cve/CVE-2026-104840
• HTTP correction: https://github.com/OpenCTI-Platform/opencti/commit/8599366319532dbdc1cfd553c55a518ad146c3df
• Login correction: https://github.com/OpenCTI-Platform/opencti/commit/d7dfab4c264e36addf215b413924cb77f0827070
• Fixed release: https://github.com/OpenCTI-Platform/opencti/releases/tag/7.260428.0
• Configuration: https://github.com/OpenCTI-Platform/opencti/blob/7.260428.0/docs/docs/deployment/configuration.md
• NVD: https://nvd.nist.gov/vuln/detail/CVE-2026-104840
• KEV: https://www.cisa.gov/known-exploited-vulnerabilities-catalog
• EPSS: https://api.first.org/data/v1/epss?cve=CVE-2026-104840
• ATT&CK: https://attack.mitre.org/techniques/T1110/001/ and https://attack.mitre.org/techniques/T1078/
• Pinned directive source: https://cdn.jsdelivr.net/npm/graphql-rate-limit-directive.0.6/dist/index.js
• Command references: https://docs.docker.com/reference/cli/docker/compose/images/ and https://docs.docker.com/reference/cli/docker/container/logs/