CVE-2025-15514: Ollama security flaw enables

CISO Take

Any organization running Ollama with multi-modal (vision) models and network-accessible API endpoints is exposed to a trivially exploitable denial-of-service — no authentication, no special knowledge, single crafted HTTP request. Patch immediately or enforce network-level access controls restricting /api/chat to trusted sources. If running Ollama in production AI pipelines or agent frameworks, treat this as a service availability incident risk until mitigated.

What is the risk?

HIGH risk for production Ollama deployments. CVSS 7.5 is accurate: network-reachable, zero authentication, low complexity, no user interaction required. The vulnerability requires only that the attacker can send an HTTP POST to /api/chat — trivially achievable against any internet-exposed instance. Ollama is among the most widely deployed local LLM runners, and many organizations expose it without authentication under the assumption it is internal. Compounding factor: vision-capable models (LLaVA, llama3.2-vision, bakllava) are increasingly used in production agentic workflows. Not in CISA KEV and no confirmed active exploitation, but the bar to exploit is extremely low and public PoC exists via the huntr advisory.

What systems are affected?

Package	Ecosystem	Vulnerable Range	Patched
Ollama	pip	—	No patch
174.6K 1.6K dependents Pushed 3d ago 12% patched ~0d to patch Full package profile →

Do you use Ollama? You're affected.

How severe is it?

CVSS 3.1

7.5 / 10

EPSS

0.6%

chance of exploitation in 30 days

Higher than 46% of all CVEs

Source: EPSS v3 — FIRST.org

Exploitation Status

Exploit Available

Exploitation: MEDIUM

Sophistication

Trivial

Exploitation Confidence

medium

○ CISA SSVC: Public PoC

○ Public PoC indexed (trickest/cve)

Composite signal derived from CISA KEV, VulnCheck KEV, CISA SSVC, EPSS, Metasploit, Exploit-DB, trickest/cve, Nuclei templates, and inthewild.io exploitation reports.

What is the attack surface?

AV Network

AC Low

PR None

UI None

S Unchanged

C None

I None

A High

What should I do?

7 steps

PATCH

Monitor ollama/ollama GitHub for a release > 0.13.5 with this fix; apply immediately when available.
NETWORK ISOLATION (immediate): Bind Ollama to localhost (default: 127.0.0.1:11434) — if exposed on 0.0.0.0, restrict via firewall rules or reverse proxy with authentication.
API GATEWAY

Place a WAF or reverse proxy in front of /api/chat that validates Content-Type and enforces payload size limits on image data fields.
INPUT VALIDATION WORKAROUND

If modifying application code, validate that base64-decoded image data has a valid MIME magic byte signature (PNG: 0x89504E47, JPEG: 0xFFD8FF) before forwarding to Ollama.
MONITORING

Alert on Ollama runner process crashes or unexpected restarts (systemd unit restart events, Docker container restarts). Alert on /api/chat requests containing images/base64 data from unexpected source IPs.
PROCESS SUPERVISION

Ensure Ollama runs with automatic restart (systemd Restart=on-failure or Docker restart policy) to minimize downtime from exploitation attempts.
AUDIT EXPOSURE

Run 'curl http://your-ollama-host:11434/api/tags' from an external network — if it responds, your instance is publicly exposed.

What does CISA's SSVC say?

Decision Track*

Exploitation poc

Automatable Yes

Technical Impact partial

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

How is it classified?

DoS Inference API Framework AML.T0029 - Denial of AI Service AML.T0034 - Cost Harvesting AML.T0040 - AI Model Inference API Access AML.T0049 - Exploit Public-Facing Application

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act

Article 15 - Accuracy, robustness and cybersecurity

ISO 42001

8.4 - AI System Operation A.6.2.6 - AI system robustness and availability

NIST AI RMF

MANAGE 2.2 - Mechanisms to sustain AI system function MANAGE-2.2 - Risk Treatment — AI System Availability

OWASP LLM Top 10

LLM04 - Model Denial of Service

Frequently Asked Questions

What is CVE-2025-15514?

Any organization running Ollama with multi-modal (vision) models and network-accessible API endpoints is exposed to a trivially exploitable denial-of-service — no authentication, no special knowledge, single crafted HTTP request. Patch immediately or enforce network-level access controls restricting /api/chat to trusted sources. If running Ollama in production AI pipelines or agent frameworks, treat this as a service availability incident risk until mitigated.

Is CVE-2025-15514 actively exploited?

Proof-of-concept exploit code is publicly available for CVE-2025-15514, increasing the risk of exploitation.

How to fix CVE-2025-15514?

1. PATCH: Monitor ollama/ollama GitHub for a release > 0.13.5 with this fix; apply immediately when available. 2. NETWORK ISOLATION (immediate): Bind Ollama to localhost (default: 127.0.0.1:11434) — if exposed on 0.0.0.0, restrict via firewall rules or reverse proxy with authentication. 3. API GATEWAY: Place a WAF or reverse proxy in front of /api/chat that validates Content-Type and enforces payload size limits on image data fields. 4. INPUT VALIDATION WORKAROUND: If modifying application code, validate that base64-decoded image data has a valid MIME magic byte signature (PNG: 0x89504E47, JPEG: 0xFFD8FF) before forwarding to Ollama. 5. MONITORING: Alert on Ollama runner process crashes or unexpected restarts (systemd unit restart events, Docker container restarts). Alert on /api/chat requests containing images/base64 data from unexpected source IPs. 6. PROCESS SUPERVISION: Ensure Ollama runs with automatic restart (systemd Restart=on-failure or Docker restart policy) to minimize downtime from exploitation attempts. 7. AUDIT EXPOSURE: Run 'curl http://your-ollama-host:11434/api/tags' from an external network — if it responds, your instance is publicly exposed.

What systems are affected by CVE-2025-15514?

This vulnerability affects the following AI/ML architecture patterns: model serving, agent frameworks, multi-modal inference pipelines, local AI development environments, RAG pipelines with image ingestion.

What is the CVSS score for CVE-2025-15514?

CVE-2025-15514 has a CVSS v3.1 base score of 7.5 (HIGH). The EPSS exploitation probability is 0.64%.

What is the AI security impact?

Affected AI Architectures

model servingagent frameworksmulti-modal inference pipelineslocal AI development environmentsRAG pipelines with image ingestion

MITRE ATLAS Techniques

AML.T0029 Denial of AI Service

AML.T0034 Cost Harvesting

AML.T0040 AI Model Inference API Access

AML.T0049 Exploit Public-Facing Application

Compliance Controls Affected

EU AI Act: Article 15

ISO 42001: 8.4, A.6.2.6

NIST AI RMF: MANAGE 2.2, MANAGE-2.2

OWASP LLM Top 10: LLM04

What are the technical details?

Original Advisory

Ollama 0.11.5-rc0 through current version 0.13.5 contain a null pointer dereference vulnerability in the multi-modal model image processing functionality. When processing base64-encoded image data via the /api/chat endpoint, the application fails to validate that the decoded data represents valid media before passing it to the mtmd_helper_bitmap_init_from_buf function. This function can return NULL for malformed input, but the code does not check this return value before dereferencing the pointer in subsequent operations. A remote attacker can exploit this by sending specially crafted base64 image data that decodes to invalid media, causing a segmentation fault and crashing the runner process. This results in a denial of service condition where the model becomes unavailable to all users until the service is restarted.

Exploitation Scenario

Attacker scans for exposed Ollama instances on port 11434 using Shodan or masscan. They identify a target running a vision model (discoverable via GET /api/tags which lists loaded models). They craft a POST request to /api/chat with model set to a multimodal model and an 'images' array containing a base64-encoded payload that decodes to invalid binary data (e.g., random bytes or a truncated file). Ollama passes this to mtmd_helper_bitmap_init_from_buf, which returns NULL due to invalid media format. The subsequent pointer dereference triggers a segfault, crashing the runner process. The attacker can repeat this every time the service restarts, maintaining a persistent DoS condition. No credentials, no AI/ML knowledge, no exploit development required — a 10-line Python script suffices. In an enterprise context where developers have enabled Ollama on their workstations for AI-assisted coding with vision capabilities, this could also be triggered by a malicious web page or document that causes the local Ollama API to be queried with crafted image data.

Weaknesses (CWE)

CWE-395 Use of NullPointerException Catch to Detect NULL Pointer Dereference Primary

CWE-395 — Use of NullPointerException Catch to Detect NULL Pointer Dereference: Catching NullPointerException should not be used as an alternative to programmatic checks to prevent dereferencing a null pointer.

[Architecture and Design, Implementation] Do not extensively rely on catching exceptions (especially for validating user input) to handle errors. Handling exceptions can decrease the performance of an application.

Source: MITRE CWE corpus.