← All security articles
HIGHAustralia (CVE-2026-125288); Customer Service Management AI agent collection < 6.1 (CVE-2026-18883)

November 2026 ServiceNow CVE Batch (KB3235180): CVE-2026-125288 template code injection and CVE-2026-18883 CSM Email response AI agent prompt injection

Domain 7: Architecture & Threat Modeling·
  • NIS2 Art.21§2(e)
  • NIS2 Art.21§2(i)
  • NIS2 Art.23§1
  • DORA Art.9§4(c)
  • DORA Art.9§4(f)
  • DORA Art.19§1
  • ISO A.8.8
  • ISO A.5.15
  • ISO A.8.2

Two CVEs, two different questions. CVE-2026-125288 affects only Australia instances below Patch 5, and the answer is your build tag. CVE-2026-18883 affects only instances with the Customer Service Management AI agent collection installed below version 6.1, and the answer is that plugin's version — your release and patch level do not decide it. ServiceNow published both on 19 November 2026 in advisory KB3235180 and reports no evidence of malicious exploitation for either.

What This Is

On 19 November 2026 ServiceNow published advisory KB3235180 with two CVEs, each found through internal testing, customer security assessments or ServiceNow's disclosure and bug-bounty programmes.

CVE CVSS v4.0 CWE Login needed? Affected Fixed in
CVE-2026-125288 (PRB2098695) 8.4 HIGH CWE-94 Yes, with template-editing rights Australia only Australia Patch 5 (also 5a, 5b W44)
CVE-2026-18883 (PRB2137625) 7.1 HIGH (none assigned) No Customer Service Management AI agent collection Plugin version 6.1

Scores, vectors and CWE assignments are as ServiceNow published them as CNA on CVE.org. KB3235180 states that CVE-2026-125288 was introduced in the Australia family release and that customers on other family releases are not impacted. It states that an instance without the CSM AI agent collection installed is not affected by CVE-2026-18883.

CVE-2026-125288 is a code injection in the ServiceNow AI Platform, ServiceNow's name for the platform itself. An authenticated user with template-editing privileges can make code run in the context of another user's session. The vector PR:H/UI:A says it plainly: the attacker needs elevated rights, and a victim has to act, by applying the template.

CVE-2026-18883 is an indirect prompt injection in the Email response AI agent of Now Assist for Customer Service Management. The Email response AI agent generates and sends an email response from source-record details [ref: markdown/intelligent-experiences/csm-email-response-ai-agent.md @ 96d32129]. Text the agent reads can carry instructions it then follows. The vector is PR:N/UI:P/VI:H: no login, passive user interaction, high integrity impact. The CVE Record assigns no CWE; CWE-1427 (improper neutralization of input used for LLM prompting) names the class. It is the only prompt-injection CVE among the 43 ServiceNow advisories the Nowisor CVE corpus tracks.

Why This Is Dangerous

Attack Scenario 1 — Session hijack through a shared template (CVE-2026-125288). An attacker holding template-editing rights, through a compromised account, an over-granted role or an insider, edits a template other users apply. When a victim applies it, the attacker's code runs as the victim. If the victim is an administrator or a fulfiller with broader rights, the attacker acts with those rights, and the audit trail records the victim, not the attacker. The precondition is the template-editor population: on many instances that population is larger than anyone has counted, because template edit rights arrive through roles granted for unrelated work.

Attack Scenario 2 — Agent steered by an inbound email (CVE-2026-18883). An attacker with no account sends text the Email response AI agent will read, for example an inbound email on a case. The text carries instructions instead of a question. On a vulnerable plugin version the agent can follow them, and its tools include drafting and sending email. An attacker can use this to make the agent send content to a recipient the attacker chooses, or with content the attacker wrote, under the instance's identity.

Attack Scenario 3 — The configuration the patch does not change. Both CVEs are fixed by updates. What the updates do not change is who can edit shared templates, and which agents act on untrusted text with no human approving the action. An instance patched against both CVEs but carrying the same over-granted template editors and the same unsupervised email-driven agents is exposed to the next flaw on the same path.

How to Detect

There is no verified detection signal for exploitation of either CVE. No indicator of compromise, payload or log pattern has been published or verified on a live instance. Detection here means establishing whether the instance is in scope, then measuring the population each flaw needs.

Determine whether CVE-2026-125288 applies

/*
 * CVE-008 - CVE-2026-125288 scope check
 * Reports the release family and patch level from the build tag.
 * Run as: Background script, admin, global scope
 * Read-only, safe for production
 */
var buildTag = gs.getProperty('glide.buildtag', '');
gs.info('Build tag: ' + buildTag);
var tag = buildTag.toLowerCase();
if (tag.indexOf('australia') < 0) {
    gs.info('Not Australia: not affected (KB3235180).');
} else {
    gs.info('Australia: affected below Patch 5.');
    gs.info('Patch 5, 5a and 5b W44 carry the fix.');
    gs.info('Compare the patch segment of the tag above.');
}

Count who can edit templates (any release)

/*
 * CVE-008 - template edit population
 * Lists active ACLs on sys_template (table and field level)
 * and the roles each one requires, then counts active
 * holders of each role. An ACL that also carries a condition
 * or script grants beyond its roles: open it in the ACL form.
 * Run as: Background script, admin, global scope
 * Read-only, safe for production
 */
var roles = {};
var acl = new GlideRecord('sys_security_acl');
acl.addEncodedQuery('active=true^name=sys_template^ORnameSTARTSWITHsys_template.');
acl.query();
while (acl.next()) {
    var op = acl.operation.getDisplayValue();
    var r = new GlideRecord('sys_security_acl_role');
    r.addQuery('sys_security_acl', acl.getUniqueValue());
    r.query();
    var names = [];
    while (r.next()) {
        var rn = r.sys_user_role.name.toString();
        names.push(rn);
        if (op == 'create' || op == 'write') {
            roles[rn] = r.getValue('sys_user_role');
        }
    }
    gs.info(acl.getValue('name') + ' ' + op + ' roles: ' + names.join(', '));
}
for (var name in roles) {
    // One row per grant path (direct, each group, inherited):
    // group by user so each person counts once.
    var h = new GlideAggregate('sys_user_has_role');
    h.addQuery('role', roles[name]);
    h.addQuery('state', 'active');
    h.addQuery('user.active', true);
    h.groupBy('user');
    h.query();
    var n = 0;
    while (h.next()) {
        n++;
    }
    gs.info('Active users holding ' + name + ': ' + n);
}

Admin holds every role, so the admin population is part of this count whatever the ACLs name.

Determine whether CVE-2026-18883 applies

/*
 * CVE-008 - CVE-2026-18883 scope check
 * Reads the installed version of the Customer Service
 * Management AI agent collection, looked up by name.
 * Run as: Background script, admin, global scope
 * Read-only, safe for production
 */
var app = new GlideRecord('sys_scope');
app.addQuery('name', 'Customer Service Management AI agent collection');
app.query();
if (!app.next()) {
    gs.info('No application found under this name.');
    gs.info('Confirm in Application Manager before concluding:');
    gs.info('KB3235180: not installed = not affected.');
} else {
    gs.info('Installed version: ' + app.getValue('version'));
    gs.info('Affected below 6.1.');
}

The name in the script is the plugin's name as ServiceNowDocs spells it [ref: markdown/customer-service-management/now-assist-for-csm/csm-live-agent-assist.md @ 96d32129]; that it is also the stored application name has not been verified on an instance. A miss therefore proves nothing on its own: confirm in Application Manager, the place KB3235180 itself points to, before concluding the plugin is absent.

Remediation

  1. Read your build tag. If it is Australia below Patch 5, apply Patch 5 or later. Hosted instances: confirm rather than assume, and raise a Now Support case quoting "November 2026 Batched CVE Release" if the build is behind. Self-hosted instances: apply the patch.

    Mitigates: Attack Scenario 1 (Session hijack through a shared template) — removes the injection path.

  2. Update the Customer Service Management AI agent collection to 6.1 or later in Application Manager, if it is installed. The fix ships in the plugin, so no platform upgrade is needed and none is sufficient.

    Mitigates: Attack Scenario 2 (Agent steered by an inbound email) — removes the known injection in the Email response AI agent.

  3. Cut the template-editor population to the people who maintain shared templates. Use the second script to list the roles the template ACLs require and who holds them. Grant edit rights on shared templates through a group that owns them, not to individuals for unrelated reasons.

    Mitigates: Attack Scenario 1 and Attack Scenario 3 — shrinks the set of accounts whose compromise turns into code running as someone else.

  4. Put a human in front of agent actions driven by customer text. For agents that read inbound email or case text and can send email or write records, set their tools to Supervised execution, which requires human oversight when the tool runs, rather than Autonomous [ref: markdown/intelligent-experiences/add-flow-action-new.md @ 96d32129].

    Mitigates: Attack Scenario 2 and Attack Scenario 3 — an injected instruction that needs a human's approval to act is a visible request, not a silent action.

Regulatory Impact

  • NIS2 Art.21§2(e) — vulnerability handling. The record expected is a dated decision per CVE: in scope or not (release family, plugin installed), and when the fix landed.
  • NIS2 Art.21§2(i) — access control. The template-editor population is the access-control fact CVE-2026-125288 depends on, and it is measurable on the instance.
  • NIS2 Art.23§1 / DORA Art.19§1 — incident notification. Both attach to an incident, not to a disclosure. ServiceNow reports no exploitation of either CVE.
  • DORA Art.9§4(f) — patches and updates. A plugin-delivered fix needs its own entry in the patch record; a platform patch record that only tracks release and patch level will not show it.
  • DORA Art.9§4(c) — logical access limited to what is needed. Applies to the template-editor roles and to the identities AI agents act as.
  • ISO A.8.8 — management of technical vulnerabilities. The build tag and the plugin version on a date are the audit artefacts.
  • ISO A.5.15 / A.8.2 — access control and privileged access rights. Template edit rights over templates others apply are a privilege in practice, whatever the role is called.

Expert Notes

  • "Not affected" means two different things here. For CVE-2026-125288 it is a release fact: Zurich and older never carried the flaw, per KB3235180. For CVE-2026-18883 it is an inventory fact: no plugin, no exposure. Neither is "past end of support, no fix published", which is what older families mean in most advisories.

  • A plugin fix does not appear in platform release notes. KB3235180 says so for PRB2137625. A patch programme that tracks only release and patch level will report the instance current while the agent collection is behind.

  • Count template editors by ACL, not by role name. The roles that grant create and write on templates are whatever the ACLs on sys_template require on your instance, plus admin. An ACL with a condition or script can grant more than its roles; a role-name search alone undercounts.

  • The patch fixes the flaw, not the pattern. CVE-2026-18883 is the first prompt-injection CVE in the Nowisor corpus, and it will not be the last on agents that read untrusted text. Supervised execution for any agent whose tools can send or write is the control that holds across the next one.