CVE-2026-65121

Executive Summary

CVE-2026-65121 is an improper authentication vulnerability (CWE-287) in the NVIDIA Infrastructure Controller for Linux, rated CVSS 8.2 (High). An attacker positioned on an adjacent network — the same broadcast domain, VLAN, or control-plane segment as the controller — can exploit the flaw without credentials or user interaction, leading to privilege escalation, information disclosure, and data tampering. Because NVIDIA Infrastructure Controller sits in the management/control plane of GPU-accelerated infrastructure (DGX/HGX systems and similarly managed nodes), successful exploitation grants an attacker influence over the administration layer rather than a single workload, which materially widens the blast radius relative to the CVSS base score alone. No public proof-of-concept, exploit code, or Nuclei template exists as of this writing, and EPSS has not yet scored the CVE.

Technical Analysis

The NVD description states that NVIDIA Infrastructure Controller for Linux "contains a vulnerability where an attacker could cause an improper authentication issue," with outcomes of privilege escalation, information disclosure, and data tampering. The assigned weakness is CWE-287 (Improper Authentication), the general class covering cases where a system either fails to verify an actor's claimed identity or verifies it incorrectly.

At the vulnerability-class level, CWE-287 manifests in control-plane software in three recurring ways: (1) an authentication check that can be skipped entirely by omitting or malforming an authentication step, (2) an authenticator (token, session identifier, signature, or credential) that is accepted without adequate validation of its properties, or (3) a trust decision made on the basis of network position, source address, or transport instead of on a cryptographic identity. The NVD text does not identify which of these applies here, and no CNA title, advisory text, or code reference is available. No specific endpoint, API path, function name, parameter, port, credential format, or protocol flow can be attributed to this CVE from the data provided, and none is fabricated in this report.

What the data does establish about the attack surface is structural:

• Adjacency is required. AV:A means the attacker must already reach the controller over an adjacent network — a shared Layer 2 segment, a management VLAN, a colocation cross-connect, or a routed segment the attacker has obtained a foothold on. This is not an internet-facing AV:N issue, but in data-center and HPC environments the "adjacent network" is frequently a flat, high-trust management subnet, so the prerequisite is often satisfied by any compromised host in the cluster.
• No privileges or interaction are needed. PR:N/UI:N indicates the authentication failure is reachable pre-authentication and requires no victim action. The vulnerable code path must therefore be reachable by an unauthenticated peer.
• Scope changes. S:C (Changed) means the vulnerable component and the impacted component are different security authorities. For a controller, the natural reading is that compromise of the controller process extends impact to the infrastructure it administers — i.e., the management plane can be used against the managed nodes.
• Confidentiality is the dominant loss (C:H), integrity is partial (I:L), availability is unaffected (A:N). This is the profile of an attacker who reads and selectively modifies protected state rather than one who destructively disrupts service.

The combination — unauthenticated, adjacency-reachable, scope-changing, confidentiality-weighted — is consistent with an authentication bypass in the controller's own management interface or inter-component trust channel, but the precise mechanism, protocol, and vulnerable code are not disclosed. Any statement more specific than this would be speculation.

Exploitation Scenarios

Note: No public PoC, exploit code, or Nuclei template is available. The following chain is inferred strictly from the vulnerability class (CWE-287) and the CVSS metrics (AV:A/AC:L/PR:N/UI:N/S:C). It contains no invented endpoints, functions, or parameters.

1) Reconnaissance and target identification. The attacker must first obtain presence on a network adjacent to the NVIDIA Infrastructure Controller for Linux. In practice this means compromising or gaining access to any host on the same management segment, VLAN, or control network, or — where management interfaces are routable — reaching the segment from a compromised workload. Target identification is by host/service discovery on that segment: identifying the controller host by its exposed management listener and service fingerprint. No CVE-specific fingerprint is published, so identification relies on knowing which assets run the product (inventory-based) rather than on a vulnerability-specific probe.

2) Triggering the flaw class. Because this is CWE-287 with PR:N, the attacker interacts with an authentication-relevant surface on the controller while not holding valid credentials. At the class level the attacker does one of the following, whichever the undisclosed flaw permits: performs the authentication exchange but omits a step the code assumes is mandatory; presents an authenticator that is not validated for the required properties (validity, binding to the request, or binding to the claimed principal); or reaches a code path that trusts the request's origin (source adjacency) in place of a verified identity. The AC:L rating indicates no race condition or unusual timing is needed and the attempt is repeatable, so a single crafted request or short request sequence is sufficient to reach the post-authentication code path.

3) Payload / execution at the vulnerability-class level. No memory corruption is implied (no A impact, no CWE-787/119 assigned). The exploitation primitive is therefore state/authorization manipulation rather than code execution: the attacker's requests are processed as though they originated from an authenticated principal. What the attacker can request once past the check is determined by the controller's management functionality, which is not documented in the available data. The attacker does not need to deliver a binary payload; the "payload" is the crafted unauthenticated request that satisfies (or circumvents) the check.

4) Post-exploitation impact. With the authentication boundary crossed, the attacker inherits the authorization context of whatever principal the vulnerable path resolves to. The CVSS subscores predict the following, and only the following: privilege escalation (the attacker operates at a level above what an unauthenticated adjacent peer should hold), information disclosure (C:H — authorized-only controller state, configuration, or managed-node information becomes readable), and data tampering (I:L — limited but real ability to alter controller-managed data or configuration). Availability is not directly impacted, so service disruption is not an expected direct outcome; the attacker's incentive is to persist quietly in the management plane, harvest data, and make targeted modifications. Because S:C, the impact is not confined to the controller process — the security authority that changes is the one the controller enforces over the infrastructure it manages.

Impact Assessment

Business impact. The affected asset is a management/control component for GPU infrastructure. Compromise of that layer is categorically more severe than compromise of a single node: controllers hold fleet-wide configuration, credentials, topology, and node state, and they are the natural pivot point for lateral movement and persistent reconfiguration across an entire accelerated-computing cluster.

Affected systems. Systems running NVIDIA Infrastructure Controller for Linux. The definite affected and fixed version ranges are not published in any tracked source, so asset owners must treat deployment inventory, not a version string, as the scoping mechanism.

Blast radius. CVSS subscores drive the assessment: AV:A narrows reachability to adjacent networks, but management networks are typically flat and high-trust, so a single compromised host on the segment is enough. PR:N/UI:N remove credential and victim-interaction barriers. S:C extends consequences beyond the vulnerable component to the assets it governs. C:H makes data exfiltration the primary risk; I:L permits targeted integrity attacks (for example, altering managed configuration) without the attacker needing to be loud; A:N means the attack is unlikely to announce itself through outage.

Severity. CVSS 8.2 High. EPSS is not yet scored, so frequency-of-exploitation cannot be estimated from FIRST.org data. CISA KEV: not listed. No public exploit exists, which currently suppresses opportunistic mass exploitation — but absence of a public PoC is a statement about tooling availability, not about attacker interest, and does not reduce the severity for environments where the management plane is reachable by untrusted or semi-trusted peers.

Detection Guidance

Detection must be behavioral, because no CVE-specific signatures, endpoints, or payload artifacts are published. The rules below target the vulnerability class: unauthenticated or anomalous authentication and authorization events against the controller, and post-authentication management activity originating from unexpected peers. Log sources assumed available: controller/service authentication logs, Linux auditd/syslog, and network flow/IDS telemetry for the management segment. Field and event names not present in the source data are generic placeholders and must be mapped to the actual product's logging schema during deployment.

Sigma Rule

title: Suspected Improper Authentication Against NVIDIA Infrastructure Controller
id: 8f2c1a44-6b7e-4d2a-9c31-0a5e7b21c9d0
status: experimental
description: >
  Detects authentication-related events against NVIDIA Infrastructure Controller for Linux
  (CVE-2026-65121, CWE-287) consistent with an unauthenticated adjacent-network actor
  reaching authenticated code paths or generating anomalous auth decisions.
references:
  - https://nvd.nist.gov/vuln/detail/CVE-2026-65121
author: valtersit.com
date: 2026/01/01
tags:
  - attack.privilege_escalation
  - attack.t1068
  - cwe.287
logsource:
  product: linux
  service: nvidia-infrastructure-controller
detection:
  selection_auth_fail:
    event_type:
      - 'authentication'
      - 'auth'
      - 'login'
      - 'session'
    outcome|contains:
      - 'failure'
      - 'denied'
      - 'invalid'
      - 'anonymous'
  selection_unexpected_principal:
    principal|contains:
      - 'anonymous'
      - 'unauthenticated'
      - 'null'
      - 'none'
  selection_privilege_context:
    event_type|contains:
      - 'privilege'
      - 'authorization'
      - 'escalat'
  condition: selection_auth_fail and (selection_unexpected_principal or selection_privilege_context)
falsepositives:
  - Legitimate service-to-service authentication misconfiguration
  - Health checks and monitoring probes against management endpoints
level: high

YARA Rule

rule CVE_2026_65121_NVIDIA_Infra_Controller_AuthBypass_Hunt
{
    meta:
        author      = "valtersit.com"
        description = "Hunts controller logs/config artifacts indicating CWE-287 authentication anomalies for CVE-2026-65121 (NVIDIA Infrastructure Controller for Linux). Generic behavioral indicators; no CVE-specific strings are published."
        reference   = "https://nvd.nist.gov/vuln/detail/CVE-2026-65121"
        cve         = "CVE-2026-65121"
        cwe         = "CWE-287"
        date        = "2026-01-01"

    strings:
        $svc1 = "Infrastructure Controller" ascii wide nocase
        $svc2 = "nvidia" ascii wide nocase

        $auth1 = "authentication" ascii wide nocase
        $auth2 = "authorization" ascii wide nocase
        $auth3 = "unauthenticated" ascii wide nocase
        $auth4 = "anonymous" ascii wide nocase

        $bad1 = "authentication bypass" ascii wide nocase
        $bad2 = "authentication skipped" ascii wide nocase
        $bad3 = "invalid session" ascii wide nocase
        $bad4 = "privilege escalation" ascii wide nocase

    condition:
        (uint32(0) != 0x7f454c46) and
        any of ($svc*) and
        any of ($auth*) and
        any of ($bad*)
}

Suricata IDS Rule

alert tcp $HOME_NET any -> $HOME_NET any (msg:"CVE-2026-65121 Possible Improper Authentication Attempt Against NVIDIA Infrastructure Controller"; flow:established,to_server; content:"Authorization"; http_header; nocase; pcre:"/Authorization\x3a\s*(?:Basic\s+Og==|Bearer\s*(?:\s|$)|(?:\r?\n))/i"; threshold:type both, track by_src, count 5, seconds 60; classtype:attempted-admin; reference:cve,2026-65121; reference:url,nvd.nist.gov/vuln/detail/CVE-2026-65121; metadata:created_at 2026_01_01, updated_at 2026_01_01, mitre_tactic_id T1068, mitre_tactic_name Privilege_Escalation; sid:202665121; rev:1;)

Elastic Detection

sequence by source.ip with maxspan=5m
  [ network where network.direction == "inbound"
      and destination.port in (443, 8443, 8080)
      and url.path != null
      and not user.name : "*" ]
  [ network where network.direction == "inbound"
      and destination.port in (443, 8443, 8080)
      and url.path != null
      and not user.name : "*"
      and http.response.status_code in (200, 201, 204) ]

Splunk SPL Query

index=infrastructure sourcetype=nvidia:controller OR sourcetype=linux:syslog
| eval evt=lower(coalesce(event_type, action, signature, "unknown"))
| eval result=lower(coalesce(outcome, status, result, "unknown"))
| where match(evt, "auth|session|login|privilege|authoriz")
| where match(result, "fail|deny|denied|invalid|anonymous|unauthent")
| bin _time span=5m
| stats count AS auth_anomalies dc(src_ip) AS distinct_sources values(dest) AS targets BY _time, src_ip
| where auth_anomalies >= 5
| sort - auth_anomalies

Wazuh OSSEC Rule

<group name="nvidia_infrastructure_controller,cve-2026-65121,">
  <rule id="202665121" level="12" frequency="3" timeframe="300">
    <if_matched_sid>202665120</if_matched_sid>
    <description>Possible improper authentication against NVIDIA Infrastructure Controller (CVE-2026-65121) - repeated authentication anomalies from same source</description>
    <mitre>
      <id>T1068</id>
    </mitre>
    <group>authentication_failures,privilege_escalation,</group>
  </rule>

  <rule id="202665120" level="5">
    <decoded_as>json</decoded_as>
    <field name="event_type">authentication|session|privilege|authorization</field>
    <field name="outcome">failure|denied|invalid|anonymous|unauthenticated</field>
    <description>Authentication anomaly on NVIDIA Infrastructure Controller (CVE-2026-65121 class)</description>
    <mitre>
      <id>T1068</id>
    </mitre>
    <group>authentication_failures,</group>
  </rule>
</group>

Mitigation & Patch Guidance

Patch confidence: 0/4 — UNPATCHED. No patch evidence exists from any tracked source. No vendor advisory, fixed version, or workaround is published; affected version ranges are also unknown. Treat all deployments of NVIDIA Infrastructure Controller for Linux as potentially affected until NVIDIA publishes authoritative guidance. Confirm current status directly with NVIDIA before acting on any specific version claim.

1) Immediate mitigations (exploit-path reduction; do not assume these fix the flaw):

• Treat the management network as hostile. Because reachability is AV:A, the highest-value immediate control is to remove adjacent-network access for anyone who should not be an administrator: strict ACLs on the management VLAN/segment, no routing of management interfaces to user or workload networks, and no shared Layer 2 with untrusted hosts.
• Segment the controller into a dedicated management enclave with allow-list-only ingress, permitting only known administrative jump hosts and required peer controller nodes.
• Require strong authentication at the network layer in front of the management interface (mutual-TLS from a hardened proxy, or an authenticating jump host), so that the vulnerable component is never directly reachable by an unauthenticated adjacent peer.
• Restrict the controller's outbound reachability to only the peers and update sources it requires, limiting lateral movement if it is compromised.
• Stand up monitoring for authentication and authorization anomalies on the management segment (see Detection Guidance) and forward controller logs off-host to a system the controller cannot reach.
• Verify — do not assume — that controller backups, configuration stores, and any cached credentials are not co-located on hosts that share the adjacent segment.

2) Patch / upgrade steps:

• There is no patch to apply. Monitor the NVIDIA security bulletin page and the NVD record for CVE-2026-65121; register for NVIDIA security notifications.
• When a fixed version is published: inventory all instances of NVIDIA Infrastructure Controller for Linux, snapshot configuration and logs, upgrade to the fixed release, verify the version post-upgrade, and re-test that the management interface is no longer reachable by unauthenticated adjacent peers.
• Until a patch exists, if the controller cannot be meaningfully isolated, consider suspending or restricting non-essential management functionality as a risk-reduction measure — accepting the operational cost deliberately, under change control.

3) Long-term hardening:

• Adopt zero-trust for infrastructure management planes: every administrative and inter-component call authenticated and authorized regardless of network origin. This directly neutralizes the CWE-287 class.
• Maintain an authoritative asset inventory with version tracking for all NVIDIA infrastructure components so that affected-range scoping is instant when advisories land.
• Enforce network microsegmentation for control-plane assets, with deny-by-default between management and workload networks and logged exceptions.
• Continuously monitor for authentication anomalies and unexpected scope-crossing events (a management-plane actor taking action against managed nodes), and review those detections after any controller software update.
• Implement credential hygiene for controller-held credentials: rotate fleet-management credentials on a schedule, and treat any period of unpatched exposure as a potential credential-compromise window requiring rotation.

References

• NVD — CVE-2026-65121: https://nvd.nist.gov/vuln/detail/CVE-2026-65121
• MITRE CWE-287 — Improper Authentication: https://cwe.mitre.org/data/definitions/287.html
• FIRST EPSS — CVE-2026-65121 (not yet scored): https://api.first.org/data/v1/epss?cve=CVE-2026-65121
• CISA Known Exploited Vulnerabilities Catalog (CVE-2026-65121 not listed): https://www.cisa.gov/known-exploited-vulnerabilities-catalog
• NVIDIA Product Security — security bulletins and advisories: https://www.nvidia.com/en-us/security/

Threat actor and exploit-in-the-wild status: No known threat actor campaigns, no CISA KEV listing, no public proof-of-concept, no exploit code, and no Nuclei template are associated with CVE-2026-65121 based on the intelligence provided. No in-the-wild exploitation evidence is available. This absence reflects current reporting as of the data snapshot, not a guarantee that exploitation will not emerge — unpatched, unauthenticated, adjacency-reachable authentication bypasses in management planes are attractive targets precisely because exploitation leaves little forensic trace and yields durable access.