CVE-2026-59225: Open WebUI arena bypass reaches

Q: Is CVE-2026-59225 actively exploited?

No confirmed active exploitation of CVE-2026-59225 has been reported, but organizations should still patch proactively.

Q: How to fix CVE-2026-59225?

Upgrade Open WebUI to 0.10.0 or later, where the task-route access check is fixed. If immediate upgrade isn't possible, treat it as a temporary compensating control: disable or restrict arena wrapper models so they don't front any model that shouldn't be accessible to the wrapper's full user base, or remove non-admin read access to arena wrappers until patched. For detection, review Open WebUI access/API logs for non-admin users calling task endpoints directly (/api/v1/tasks/moa/completions and similar) rather than the standard chat completion route, and cross-check which underlying models those calls resolved to against each user's actual model permissions.

Q: What systems are affected by CVE-2026-59225?

This vulnerability affects the following AI/ML architecture patterns: model serving, agent frameworks.

Q: What is the CVSS score for CVE-2026-59225?

CVE-2026-59225 has a CVSS v3.1 base score of 5.4 (MEDIUM). The EPSS exploitation probability is 0.21%.

CISO Take

Open WebUI's arena feature lets non-admin users compare LLM outputs through a wrapper model, but a flaw in versions 0.8.12 through 0.10.0 lets any authenticated user with read access to that wrapper reach the underlying restricted model directly, by calling task endpoints such as /api/v1/tasks/moa/completions instead of the normal chat route — those endpoints call the completion generator directly and resolve the arena fallback after the wrapper access check, then recurse with bypass_filter=True and skip the submodel's own check. The blast radius today is limited: no downstream package dependents are recorded, there's no public exploit or Nuclei template, it isn't in CISA KEV, and EPSS isn't available, but the bug needs only low privileges and no user interaction (CVSS 5.4, AC:L), so any org relying on Open WebUI model tiers to gate expensive, fine-tuned, or compliance-restricted models from general users cannot trust that boundary as configured. Open-webui also carries 130 other recorded CVEs and a package risk score of 38/100, which argues for treating its access-control layer as generally fragile rather than a one-off. Patch to 0.10.0 now; until then, don't expose any model via an arena wrapper that you wouldn't also grant directly, and audit logs for non-admin calls to /api/v1/tasks/* as a detection signal for exploitation attempts.

Sources: NVD GitHub Advisory CISA KEV ATLAS

What is the risk?

Medium overall risk. Exploitability is high on the technical axis (network-reachable, low attack complexity, no user interaction) but requires an authenticated account with at least low privileges and specific pre-existing read access to an arena wrapper model, which narrows real-world exposure to orgs that actively use arena/MoA routing. Impact is bounded — confidentiality-low, integrity-none, availability-low per the CVSS vector — because the attacker gains unauthorized inference access to a model, not code execution or data destruction. No KEV listing, no EPSS score, no public PoC, and no scanner template mean opportunistic mass exploitation is unlikely; this is a targeted-misuse risk (a disgruntled or curious internal user, or a compromised low-privilege account) rather than an internet-scanning risk.

How does the attack unfold?

Initial Access

Attacker authenticates as a low-privileged, non-admin Open WebUI user who has been granted read access to an arena wrapper model.

AML.T0012

Endpoint Abuse

Attacker bypasses the normal chat route and calls a task endpoint directly, such as /api/v1/tasks/moa/completions, which does not re-check the resolved submodel's access permissions.

AML.T0047

Authorization Bypass

The arena wrapper resolves to its underlying restricted submodel and the task route recurses with bypass_filter=True, skipping the submodel's own access check.

AML.T0040

Impact

Attacker obtains unauthorized inference access to a model they were never directly granted, undermining model-tier access restrictions and any compliance controls built on them.

AML.T0040

Initial Access

Attacker authenticates as a low-privileged, non-admin Open WebUI user who has been granted read access to an arena wrapper model.

AML.T0012

Endpoint Abuse

Attacker bypasses the normal chat route and calls a task endpoint directly, such as /api/v1/tasks/moa/completions, which does not re-check the resolved submodel's access permissions.

AML.T0047

Authorization Bypass

The arena wrapper resolves to its underlying restricted submodel and the task route recurses with bypass_filter=True, skipping the submodel's own access check.

AML.T0040

Impact

Attacker obtains unauthorized inference access to a model they were never directly granted, undermining model-tier access restrictions and any compliance controls built on them.

AML.T0040

What systems are affected?

Package	Ecosystem	Vulnerable Range	Patched
Open WebUI	pip	>= 0.8.12, < 0.10.0	`0.10.0`
149.6K 4 dependents Pushed 3d ago 81% patched ~6d to patch Full package profile →

Do you use Open WebUI? You're affected.

How severe is it?

CVSS 3.1

5.4 / 10

EPSS

0.2%

chance of exploitation in 30 days

Higher than 11% of all CVEs

Source: EPSS v3 — FIRST.org

Exploitation Status

No known exploitation

Sophistication

Moderate

What is the attack surface?

AV Network

AC Low

PR Low

UI None

S Unchanged

C Low

I None

A Low

What should I do?

1 step

Upgrade Open WebUI to 0.10.0 or later, where the task-route access check is fixed. If immediate upgrade isn't possible, treat it as a temporary compensating control: disable or restrict arena wrapper models so they don't front any model that shouldn't be accessible to the wrapper's full user base, or remove non-admin read access to arena wrappers until patched. For detection, review Open WebUI access/API logs for non-admin users calling task endpoints directly (/api/v1/tasks/moa/completions and similar) rather than the standard chat completion route, and cross-check which underlying models those calls resolved to against each user's actual model permissions.

What does CISA's SSVC say?

Decision Track

Exploitation none

Automatable No

Technical Impact partial

Source: CISA Vulnrichment (SSVC v2.0). Decision based on the CISA Coordinator decision tree.

How is it classified?

Auth Bypass Model Inference AML.T0012 - Valid Accounts AML.T0040 - AI Model Inference API Access AML.T0047 - AI-Enabled Product or Service

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act

Article 15 - Accuracy, robustness and cybersecurity

ISO 42001

A.6.2.2 - AI system operational access controls

OWASP LLM Top 10

LLM06 - Excessive Agency

Frequently Asked Questions

What is CVE-2026-59225?

Open WebUI's arena feature lets non-admin users compare LLM outputs through a wrapper model, but a flaw in versions 0.8.12 through 0.10.0 lets any authenticated user with read access to that wrapper reach the underlying restricted model directly, by calling task endpoints such as /api/v1/tasks/moa/completions instead of the normal chat route — those endpoints call the completion generator directly and resolve the arena fallback after the wrapper access check, then recurse with bypass_filter=True and skip the submodel's own check. The blast radius today is limited: no downstream package dependents are recorded, there's no public exploit or Nuclei template, it isn't in CISA KEV, and EPSS isn't available, but the bug needs only low privileges and no user interaction (CVSS 5.4, AC:L), so any org relying on Open WebUI model tiers to gate expensive, fine-tuned, or compliance-restricted models from general users cannot trust that boundary as configured. Open-webui also carries 130 other recorded CVEs and a package risk score of 38/100, which argues for treating its access-control layer as generally fragile rather than a one-off. Patch to 0.10.0 now; until then, don't expose any model via an arena wrapper that you wouldn't also grant directly, and audit logs for non-admin calls to /api/v1/tasks/* as a detection signal for exploitation attempts.

Is CVE-2026-59225 actively exploited?

No confirmed active exploitation of CVE-2026-59225 has been reported, but organizations should still patch proactively.

How to fix CVE-2026-59225?

Upgrade Open WebUI to 0.10.0 or later, where the task-route access check is fixed. If immediate upgrade isn't possible, treat it as a temporary compensating control: disable or restrict arena wrapper models so they don't front any model that shouldn't be accessible to the wrapper's full user base, or remove non-admin read access to arena wrappers until patched. For detection, review Open WebUI access/API logs for non-admin users calling task endpoints directly (/api/v1/tasks/moa/completions and similar) rather than the standard chat completion route, and cross-check which underlying models those calls resolved to against each user's actual model permissions.

What systems are affected by CVE-2026-59225?

This vulnerability affects the following AI/ML architecture patterns: model serving, agent frameworks.

What is the CVSS score for CVE-2026-59225?

CVE-2026-59225 has a CVSS v3.1 base score of 5.4 (MEDIUM). The EPSS exploitation probability is 0.21%.

What is the AI security impact?

Affected AI Architectures

model servingagent frameworks

MITRE ATLAS Techniques

AML.T0012 Valid Accounts

AML.T0040 AI Model Inference API Access

AML.T0047 AI-Enabled Product or Service

Compliance Controls Affected

EU AI Act: Article 15

ISO 42001: A.6.2.2

OWASP LLM Top 10: LLM06

What are the technical details?

Original Advisory

Open WebUI is an extensible, feature-rich, and user-friendly self-hosted AI platform. From 0.8.12 before 0.10.0, an authenticated non-admin user with read access to an arena wrapper model can reach a restricted underlying model through task endpoints such as /api/v1/tasks/moa/completions. The normal chat route resolves arena models before the final chat dispatch and therefore re-checks the selected underlying model. The task routes call utils.chat.generate_chat_completion() directly. In that direct path, arena fallback resolution happens after the wrapper access check and then recurses with bypass_filter=True, skipping the selected submodel's access check. This issue is fixed in version 0.10.0.

Exploitation Scenario

A marketing analyst has been granted read access to an arena wrapper model used to A/B test chatbot responses, but has not been granted access to the high-cost frontier model that the wrapper sometimes routes to internally. Instead of using the normal chat UI, the analyst (or a script acting on their session) calls the task API directly — POST /api/v1/tasks/moa/completions — specifying the arena wrapper as the target. Open WebUI's task-route logic resolves the arena fallback after the initial wrapper-level access check, then recurses into the underlying restricted model with bypass_filter=True, skipping the check that would normally block the analyst from that submodel. The analyst now gets direct completions from a model they were explicitly denied, at no additional cost visibility or audit trail beyond raw API logs.

Weaknesses (CWE)

CWE-862 Missing Authorization Primary CWE-862 Missing Authorization

CWE-862 — Missing Authorization: The product does not perform an authorization check when an actor attempts to access a resource or perform an action.

[Architecture and Design] Divide the product into anonymous, normal, privileged, and administrative areas. Reduce the attack surface by carefully mapping roles with data and functionality. Use role-based access control (RBAC) [REF-229] to enforce the roles at the appropriate boundaries. Note that this approach may not protect against horizontal authorization, i.e., it will not protect a user from attacking others with the same role.
[Architecture and Design] Ensure that access control checks are performed related to the business logic. These checks may be different than the access control checks that are applied to more generic resources such as files, connections, processes, memory, and database records. For example, a database may restrict access for medical records to a specific database user, but each record might only be intended to be accessible to the patient and the patient's doctor [REF-7].

Source: MITRE CWE corpus.