CVE-2026-41523

EUVD-2026-38406 · Published Jun 16, 2026

vLLM: Security Check Bypass via assert Statement in Activation Function Loading Allows Arbitrary Code Execution

Summary

An assert-based security check in vLLM's activation function loading allows any unauthenticated attacker to achieve arbitrary code execution on the server by publishing a malicious HuggingFace model, when vLLM runs in Python optimized mode (python -O or PYTHONOPTIMIZE=1).

Details

vLLM uses an assert statement at vllm/model_executor/layers/pooler/activations.py:48 as its sole security control to restrict which activation functions can be loaded from a HuggingFace model's config.json:

# vllm/model_executor/layers/pooler/activations.py:35-53
function_name: str | None = None
if (
    hasattr(config, "sentence_transformers")
    and "activation_fn" in config.sentence_transformers
):
    function_name = config.sentence_transformers["activation_fn"]
elif (
    hasattr(config, "sbert_ce_default_activation_function")
    and config.sbert_ce_default_activation_function is not None
):
    function_name = config.sbert_ce_default_activation_function

if function_name is not None:
    assert function_name.startswith("torch.nn.modules."), (
        "Loading of activation functions is restricted to "
        "torch.nn.modules for security reasons"
    )
    fn = resolve_obj_by_qualname(function_name)()

Python's assert statements are stripped at compile time when running in optimized mode (python -O or PYTHONOPTIMIZE=1). When the assert is absent, the attacker-controlled function_name from the model's config.json is passed directly to resolve_obj_by_qualname() — an unrestricted import gadget:

def resolve_obj_by_qualname(qualname: str) -> Any:
    module_name, obj_name = qualname.rsplit(".", 1)
    module = importlib.import_module(module_name)
    return getattr(module, obj_name)

This is the same vulnerability class as CVE-2017-1000433 (pysaml2 assert-based auth bypass), flagged by Bandit B101 and Ruff S101, and the reason Django proactively replaced all assert-based security checks (ticket #32508).

Attacker-controlled input sources:

config.sentence_transformers["activation_fn"] (line 40)
config.sbert_ce_default_activation_function (line 45)

Affected call sites — get_act_fn() is called via resolve_classifier_act_fn() from:

vllm/model_executor/layers/pooler/seqwise/poolers.py:122 — SequencePooler
vllm/model_executor/layers/pooler/tokwise/poolers.py:130 — TokenPooler

Broader systemic risk: resolve_obj_by_qualname is called from ~20 locations across the codebase with no validation of its own. Any future caller feeding user-controlled input to it without validation creates the same vulnerability class.

Suggested fix: Replace the assert with an explicit conditional raise:

if not function_name.startswith("torch.nn.modules."):
    raise ValueError(
        "Loading of activation functions is restricted to "
        "torch.nn.modules for security reasons"
    )

Impact

Arbitrary code execution. A malicious model author publishes a HuggingFace model with a crafted config.json. When a victim loads this model with vLLM running under python -O or PYTHONOPTIMIZE=1, arbitrary code executes during model initialization with the privileges of the vLLM process.

The attack requires:

Victim loads a malicious model from HuggingFace (user interaction)
vLLM runs under python -O or PYTHONOPTIMIZE=1 (documented in production use)
Model uses a cross-encoder architecture (e.g. BERT or RoBERTa with sequence classification)

Coordinated disclosure note: This vulnerability was also reported via huntr.com on April 2, 2026 (https://huntr.com/bounties/dcb05b04-e625-41e7-adbc-bbae0cc2d64c). A GitHub Security Advisory was also filed because it is vLLM's stated preferred disclosure channel per SECURITY.md.

Fix

A fix for this was introduced in this commit: https://github.com/vllm-project/vllm/commit/b3c7ffcab82c2439726f8cb213800f6f38c023d3

Technologies

Vllm Vllm

Weaknesses

CWE-94CWE-617

CVSS scores

v3.17.5CVSS:3.1/AV:N/AC:H/PR:N/UI:R/S:U/C:H/I:H/A:H

MITRE ATT&CK

Inferred

Where this vulnerability fits in an attack

ReconImpact

Initial Access Stealth

T1190Exploit Public-Facing Application T1027.006HTML Smuggling T1027.009Embedded Payloads T1564.009Resource Forking

References

Potentially impacted assets

See if this affects your attack surface

Latest trending attack

Criticalyesterday

Google Brotli decoder - Remote Code Execution via compressed requests (Exploitation Check)

This control is the second step of the Patrowl response to CVE-2026-111497 (Node2Shell). It complements the first control of 22 October 2026, which focused on the Node.js runtime, and does not replace it. Why a second control CVE-2026-111497 is an out-of-bounds write in the Google Brotli decoder, the library that decompresses Brotli ("br") data. It was first published as a Node.js vulnerability, because Node.js bundles this decoder. Later publications showed that the flaw is in the library itself. Every product that embeds an affected version of the decoder is concerned, whatever its language or vendor. A public proof of concept demonstrates code execution. Any server that decompresses request bodies sent with "Content-Encoding: br" exposes the decoder to unauthenticated attackers. Known cases include Express 5 and NestJS 11 on Node.js, Koa and Fastify setups, Bun, Envoy and Istio with request decompression enabled, and any application or reverse proxy that decodes Brotli request bodies with an affected library. The flaw also works in the other direction. A malicious server can compromise an unpatched client that decodes a Brotli response, for example Node.js fetch(), Python clients (requests, httpx, aiohttp, Scrapy) with the Brotli or brotlicffi package, or libcurl builds with Brotli support. What this control checks Patrowl sends small, valid Brotli-compressed requests to each asset and observes whether the server decodes them. The check does not depend on the technology: it detects the behaviour, whatever the stack behind it. These requests are what any legitimate client could send. They never reach the vulnerable code path and cannot affect the availability of the service. How to read the result Potentially impacted means the asset decodes Brotli request bodies: the attack surface is exposed. It does not confirm that the asset is vulnerable. The version of the Brotli decoder in use is not visible from the outside. This check covers applications that parse request bodies on their root path. A Not impacted result does not rule out exposure on other routes, nor on the client side. Checking the versions on the servers remains the only exhaustive verification. Two situations are possible. When the decoding is done by your own application or reverse proxy, identify the component and update it. If an update is not possible right away, reject requests with "Content-Encoding: br" at the reverse proxy or WAF. When the asset is hosted on a managed platform or behind a CDN (for example Platform.sh, Upsun or Fastly), the decoding may be done by the provider's edge router, in front of your application. You cannot patch it yourself, but this router handles all the traffic of your site. Ask your provider to confirm that its edge components are patched against CVE-2026-111497. Assets hosted on a managed platform or behind a CDN carry the tag "paas: platform.sh" or "cdn: fastly" in the asset view: for these assets, the Brotli decoding is done by the provider's edge router, in front of your application, and only the provider can patch it. Affected and fixed versions Google Brotli 1.0.3 to 1.2.0 are affected, fixed in 1.0.10, 1.1.1 and 1.2.1. Fixed releases of the main products embedding it: Node.js 22.23.4, 24.22.1 and 26.12.1 (end-of-life lines 20.x and older and 25.x receive no fix), Bun 1.4.4, Envoy 1.37.8, 1.38.6, 1.39.3 and 1.40.0, Istio 1.29.9, 1.30.6 and 1.31.2, Python Brotli 1.2.1 and brotlicffi 1.2.1.0. Products that link against the system library depend on the Brotli package of the Linux distribution. Recommendations Update every component that embeds the Brotli decoder (only real fix). Start with the assets flagged by this control. Restart the processes after the update. Rebuild containers, single-executable applications and appliances that bundle an affected runtime or library. Reduce the exposure until the update is deployed. Reject requests with "Content-Encoding: br" at the reverse proxy, load balancer or WAF, case-insensitively and including lists such as "gzip, br". On Envoy and Istio, disable request decompression for Brotli. Contact your edge provider about the status of the pathing (Fastly / Plateform.sh) Review outbound calls. List the features that fetch content from servers you do not control, and patch the services that host them first.