Key facts
| Fact | Detail |
|---|---|
| Product | Ollama (AI model server) |
| Affected versions | 0.34.2 and later, below 0.35.0; older versions: status unknown |
| Fixed in | 0.35.0 and later (latest release: 0.40.10) |
| Severity | 9.4 Critical (CVSS 4.0, CERT Polska) |
| Who can attack | Anyone who can connect to port 11434 and run a model registry your server can reach; no login needed |
| Exploited | No attacks reported as of 23 October 2026 (not in CISA's KEV catalog); public attack code since 8 October 2026; ready-to-run kit on GitHub since 19 October 2026 |
| CVE published | 8 October 2026 |
What an attacker can do
Ollama runs AI models on your own servers. Its HTTP API listens on port 11434 and has no login.
The flaw is a path traversal (a file name with ../ that escapes its folder) in model downloads (POST /api/pull). Ollama reads a model's file list from a registry (a server that hosts models) and does not check the file names. So an attacker's registry can make your server write files outside its model folder.
On the default Docker image, the attacker plants a malicious library file in /usr/lib/ollama, which Ollama loads at the next restart. The attacker can then run any command as root inside the container (an isolated space on the server). CERT Polska, the Polish security team that registered this flaw, rates the impact on confidentiality, integrity and availability as High.
Who is affected
The CVE record lists Ollama 0.34.2 and later, below 0.35.0.
- Official Docker image: Ollama runs as root and can write to
/usr/lib/ollama. It listens on every network interface of the container (OLLAMA_HOST=0.0.0.0:11434). With the documented-p 11434:11434, Docker opens that port on all the server's networks. CERT Polska scores this case 9.4 Critical (CVSS 4.0, the standard severity scale out of 10). - Standard Linux install with
install.sh: the service runs as the userollamaand does not reach root this way. The attacker can still create files outside the model folder, or wipe the content of existing ones. This case scores 6.9 Medium.
The attacker needs access to port 11434, and your server must reach a registry the attacker controls (the public attack code uses plain HTTP). A native install listens on 127.0.0.1 (the machine itself) by default. On some servers with a GPU, the planted code runs without a restart.
How to check your version
Run one of these commands. On the server that runs Ollama:
ollama -v
Over the API, replace localhost with your server address:
curl http://localhost:11434/api/version
The answer looks like {"version": "..."}.
On a Docker install, replace ollama with your container name:
docker exec -it ollama ollama -v
Your server is vulnerable if the version is 0.34.2 or later and below 0.35.0. Update older versions too, because the CVE record gives them the status "unknown" and does not declare them safe.
Ollama published no advisory or workaround of its own. The notes of the patched release do not mention the flaw either, so update without waiting for a security bulletin.
What to do now
Update Ollama to 0.35.0 or later. On Linux, run the install script again:
curl -fsSL https://ollama.com/install.sh | sh
On macOS and Windows, the app downloads updates: click "Restart to update". On Docker, download the image ollama/ollama:0.35.0 or newer and recreate the container.
Keep port 11434 off the internet and off untrusted networks. For remote access, put a proxy server that asks for a login in front of the API.
Neither Ollama nor Striga, the company that found the flaw, listed signs of compromise, but the company's write-up (dated 8 October 2026) and attack kit (on GitHub since 19 October 2026) give two leads:
- an
mlx_*folder or alibmlxc.*file in/usr/lib/ollamathat the official image does not ship (the attack code plantsmlx_zzz/libmlxc.tmp); - model pulls from unknown registries with
"insecure":true, the plain HTTP option, in request bodies captured by a reverse proxy or a network capture. At the default log level, Ollama's own log shows onlyPOST "/api/pull", without the model name or the insecure flag. WithOLLAMA_DEBUG=1set, look for adownloading blobline whose digest contains../or does not start withsha256.
Get help
Updating closes this flaw. Flaws in your own code get no CVE and no update, and a test of your app is how they are found.
AI agent and chatbot security audit
Can a message, an email or a document make your AI agent act against you? We try, and show you how to stop it.
From AED 35,000, excl. 5% VAT
Sources
- CVE-2026-103663 record CVE Program
- Vulnerability in Ollama software (CVE-2026-103663) CERT Polska
- DirtyDigest: A crafted manifest turns Ollama's model pull into an arbitrary file write and root RCE Striga (finder)
- striga-ai/CVE-2026-103663 reproduction kit Striga (finder)
- Ollama v0.35.0 release Ollama
- Known Exploited Vulnerabilities Catalog CISA