high CVSS 8.8

CVE-2026-22807·Python vulnerability

vLLM is an inference and serving engine for large language models (LLMs). Starting in version 0.10.1 and prior to version 0.14.0, vLLM loads Hugging Face `auto_map` dynamic modules during model resolution without gating on `trust_remote_code`, allowing attacker-controlled Python code in a model repo/path to execute at server startup. An attacker who can influence the model repo/path (local directory or remote Hugging Face repo) can achieve arbitrary code execution on the vLLM host during model load. This happens before any request handling and does not require API access. Version 0.14.0 fixes the issue.

Severity
high
Software
Python
Fixed in
0.14.0
Published
2026-01-21

Affected versions

From: 0.10.1

Until: 0.14.0

Fixed in: 0.14.0

How to fix this CVE

Update vLLM to version 0.14.0 or later to prevent arbitrary code execution during model loading. The vulnerability allows untrusted Python code from Hugging Face model repositories to execute at server startup without requiring the `trust_remote_code` flag. Immediately patch all vLLM instances running versions 0.10.1 through 0.13.x, and review any model repositories used in your environment for potentially malicious code.

sudo dnf update python3-vllm || sudo pip3 install --upgrade vllm>=0.14.0

Defensia detects this vulnerability

How to check if you are affected

  1. Check vLLM version: pip3 show vllm | grep Version or python3 -c "import vllm; print(vllm.__version__)"
  2. Identify model loading paths: grep -r 'auto_map\|trust_remote_code' /path/to/vllm/config or examine model configuration files for remote Hugging Face repositories
  3. Review server startup logs for unexpected Python module imports or execution: journalctl -u vllm -n 500 | grep -i 'import\|module\|load\|exec' or check /var/log/vllm/* for anomalies
  4. Verify patched version is running: pip3 show vllm | grep -E 'Version.*0\.1[4-9]|Version.*0\.[2-9]' to confirm 0.14.0 or later is installed

FAQ

What is CVE-2026-22807?

CVE-2026-22807 is a critical vulnerability in vLLM versions 0.10.1 to 0.13.x where the dynamic module loading mechanism bypasses security controls, allowing arbitrary Python code execution from untrusted model repositories during server startup without user consent.

Is CVE-2026-22807 being actively exploited?

No, this vulnerability is not currently listed on the CISA Known Exploited Vulnerabilities (KEV) catalog and no public exploits are available, but the attack vector is straightforward and organizations should patch proactively.

What versions of vLLM are affected by CVE-2026-22807?

vLLM versions 0.10.1 through 0.13.x are vulnerable. Version 0.14.0 and later include the fix that gates auto_map module loading on the `trust_remote_code` parameter.

How do I check if my server is vulnerable to CVE-2026-22807?

Run `pip3 show vllm | grep Version` and verify the version is less than 0.14.0; if so, your deployment is vulnerable. Also confirm that model loading configurations do not explicitly enable remote code execution guards.

Does Defensia detect CVE-2026-22807?

Yes — Defensia's CVE advisory scanner compares installed package versions against the NVD database. If vLLM is installed on a monitored server, CVE-2026-22807 will appear in your dashboard with remediation steps.

Related Python CVEs

CVE-2026-34938CVSS 10PraisonAI is a multi-agent teams system. Prior to version 1.5.90, execute_code() in praisonai-agents runs attacker-controlled Python inside a three-layer sandbox that can be fully bypassed by passing a str subclass with an overridden startswith() method to the _safe_getattr wrapper, achieving arbitrary OS command execution on the host. This issue has been patched in version 1.5.90.
CVE-2026-33054CVSS 10Mesop is a Python-based UI framework that allows users to build web applications. Versions 1.2.2 and below contain a Path Traversal vulnerability that allows any user supplying an untrusted state_token through the UI stream payload to arbitrarily target files on the disk under the standard file-based runtime backend. This can result in application denial of service (via crash loops when reading non-msgpack target files as configurations), or arbitrary file manipulation. This vulnerability heavily exposes systems hosted utilizing FileStateSessionBackend. Unauthorized malicious actors could interact with arbitrary payloads overwriting or explicitly removing underlying service resources natively outside the application bounds. This issue has been fixed in version 1.2.3.
CVE-2026-28505CVSS 10Tautulli is a Python based monitoring and tracking tool for Plex Media Server. Prior to version 2.17.0, the str_eval() function in notification_handler.py implements a sandboxed eval() for notification text templates. The sandbox attempts to restrict callable names by inspecting code.co_names of the compiled code object. However, co_names only contains names from the outer code object. When a lambda expression is used, it creates a nested code object whose attribute accesses are stored in code.co_consts, NOT in code.co_names. The sandbox never inspects nested code objects. This issue has been patched in version 2.17.0.
CVE-2026-25632CVSS 10EPyT-Flow is a Python package designed for the easy generation of hydraulic and water quality scenario data of water distribution networks. Prior to 0.16.1, EPyT-Flow’s REST API parses attacker-controlled JSON request bodies using a custom deserializer (my_load_from_json) that supports a type field. When type is present, the deserializer dynamically imports an attacker-specified module/class and instantiates it with attacker-supplied arguments. This allows invoking dangerous classes such as subprocess.Popen, which can lead to OS command execution during JSON parsing. This also affects the loading of JSON files. This vulnerability is fixed in 0.16.1.
CVE-2026-8476CVSS 9.9IBM Langflow OSS 1.0.0 through 1.10.0 contain a critical remote code execution vulnerability in the disk-based caching mechanism. The AsyncDiskCache class uses Python's unsafe pickle.loads() function to deserialize cached objects from disk without validation, integrity verification, or authentication, enabling arbitrary code execution when malicious pickle payloads are processed. Attackers who can influence cached data through file system access, malicious workflow inputs, custom components, or API manipulation can achieve complete system compromise with the privileges of the Langflow server process.

References

Track CVEs across your fleet automatically

Defensia scans your Linux servers and tells you exactly which ones are running vulnerable versions — including CVE-2026-22807. Free for 1 server.

Get started free