CVE-2026-90970

validatedZero produced an exploit or a check.

GHSA-5295-vp56-jghq ↗EUVD-2026-91651 ↗

check availableZero validated a working check for it.not proven exploitableZero has a detection check, but hasn't produced a full working exploit yet.
cve.orgpublic pocExploit code has been published by someone. Source signals can also carry a PoC (Proof of Concept).kevAttacks confirmed in the real world.
GitLab AI Gateway >= 18.1.6, < 19.2.4; >= 19.3, < 19.3.2; >= 19.4, < 19.4.1
fixed in 19.2.4, 19.3.2, 19.4.1
CWE-1336 Improper Neutralization of Special Elements Used in a Template Engine
ETK-2026-346, permanent identifier
+37mTTV
41mAI analysisAnalysis: how long the AI spent reversing this signal.
Check if you are vulnerable
timeline
0 = trigger signal
trigger signal (cve.org) +0mcheck available+37magent stopped
0
+10min
+20min
+30min
+40min

description

GitLab AI Gateway versions from 18.1.6 before 19.2.4, 19.3 before 19.3.2, and 19.4 before 19.4.1 are affected. An authenticated user with Duo Agent Platform access may be able to execute commands on the AI Gateway through a malicious flow configuration.

mitigation

Upgrade GitLab AI Gateway to 19.2.4, 19.3.2, or 19.4.1, as appropriate for the deployed release line. Until upgraded, restrict Duo Agent Platform access to trusted users.

fixed in 19.2.4, 19.3.2, 19.4.1

check
available
exploit
no
automationDeterministic: a check that runs on its own, the same request on every target. Contextual: an agentic technique that needs to be applied per target, for bugs that depend on contextual flows.
contextual
auth requiredWhether an attacker needs an account first, based on the CVSS vector's Privileges Required.
yes, low privileges
EPSSEstimate of the chance this CVE is exploited in the next 30 days. Refreshed daily.
0.94%60th percentile
agent run time
41m
cve.org record, t = 0The moment cve.org publishes the CVE in full.
2026-10-02 14:34Z
Δ to public pocExploit code has been published by someone. Source signals can also carry a PoC (Proof of Concept).
–
Δ to kevAttacks confirmed in the real world.
–

outcome

Contextual check: a technique exists and was validated, but no deterministic check could cover this one.

source signals

Public signals that triggered or updated the pipeline. These are inputs for the deltas above.

  1. 2026-10-02 14:34Z
    cve_org: trigger signal, started the run
  2. 2026-10-02 17:33Z
    GitLab Patches Critical 9.9 AI Gateway Flaw Allowing Command Execution on Self-Hosted Servers
  3. 2026-10-13 08:43Z
    GitLab Patches CVSS 10.0 Flaw That Could Let Unauthenticated Users Run Server Code

this entry in the feed

GET /api/v1/feed/CVE-2026-90970

{
  "etk_id": "ETK-2026-346",
  "cve_id": "CVE-2026-90970",
  "display_id": "CVE-2026-90970",
  "ghsa_id": "GHSA-5295-vp56-jghq",
  "euvd_id": "EUVD-2026-91651",
  "status": "published",
  "anchor_state": "anchored",
  "title": "Prompt-template sandbox escape in GitLab AI Gateway",
  "summary": "GitLab AI Gateway versions from 18.1.6 before 19.2.4, 19.3 before 19.3.2, and 19.4 before 19.4.1 are affected. An authenticated user with Duo Agent Platform access may be able to execute commands on the AI Gateway through a malicious flow configuration.",
  "mitigation": "Upgrade GitLab AI Gateway to 19.2.4, 19.3.2, or 19.4.1, as appropriate for the deployed release line. Until upgraded, restrict Duo Agent Platform access to trusted users.",
  "vendor": "GitLab",
  "product": "GitLab AI Gateway",
  "affected": [
    {
      "affected_notes": [],
      "affected_ranges": [
        ">= 18.1.6, < 19.2.4",
        ">= 19.3, < 19.3.2",
        ">= 19.4, < 19.4.1"
      ],
      "ecosystem": null,
      "fixed_versions": [
        "19.2.4",
        "19.3.2",
        "19.4.1"
      ],
      "package_name": "GitLab AI Gateway",
      "purl": null
    }
  ],
  "cwes": [
    1336
  ],
  "cvss_score": 9.9,
  "cvss_src": "cna_vendor",
  "cvss_vector": "CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:C/C:H/I:H/A:H",
  "epss_score": 0.0094,
  "epss_percentile": 0.6044,
  "epss_asof": "2026-10-20",
  "automation": "contextual",
  "check": true,
  "check_confidence_basis": "assessed",
  "exploit": false,
  "pipeline_minutes": 40.5913498,
  "t0_at": "2026-10-02T14:34:49.558000+00:00",
  "signals": {
    "check": {
      "at": "2026-10-02T15:12:01.618015+00:00",
      "days": 0.02583402795138889,
      "is_lower_bound": false,
      "measured_at": "2026-10-02T14:34:49.558000+00:00",
      "url": null
    },
    "trigger_signal": {
      "at": "2026-10-02T14:34:49.558000+00:00",
      "days": 0,
      "is_lower_bound": false,
      "measured_at": "2026-10-02T14:34:49.558000+00:00",
      "url": "https://www.cve.org/CVERecord?id=CVE-2026-90970"
    }
  }
}

published 2026-10-02 15:01Z
last change 2026-10-21 10:52Z

Need the technical details behind an entry?

Every entry here shows what we attempted and when. The technical details, the generated exploit that fired it and the evidence behind the verdict are not public. Tell us about your use case and we’ll get back to you about extended API access.