Data exposure via ZeroMQ on multi-node vLLM deployment
vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs. Versions starting from 0.5.2 and prior to 0.8.5 are vulnerable to denial of service and data exposure via ZeroMQ on multi-node vLLM deployment. In a multi-node vLLM deployment, vLLM uses ZeroMQ for some multi-node communication purposes. The primary vLLM host opens an XPUB ZeroMQ socket and binds it to ALL interfaces. While the socket is always opened for a multi-node deployment, it is only used when doing tensor parallelism across multiple hosts. Any client with network access to this host can connect to this XPUB socket unless its port is blocked by a fire
Regular Expression Denial of Service (ReDoS) in huggingface/transformers
A Regular Expression Denial of Service (ReDoS) vulnerability was identified in the huggingface/transformers library, specifically in the file `tokenization_gpt_neox_japanese.py` of the GPT-NeoX-Japanese model. The vulnerability occurs in the SubWordJapaneseTokenizer class, where regular expressions process specially crafted inputs. The issue stems from a regex exhibiting exponential complexity under certain conditions, leading to excessive backtracking. This can result in high CPU usage and potential application downtime, effectively creating a Denial of Service (DoS) scenario. The affected version is v4.48.1 (latest).
inclusionAI AWorld shell_tool.py subprocess.Popen os command injection
A vulnerability was found in inclusionAI AWorld up to 8c257626e648d98d793dd9a1a950c2af4dd84c4e. It has been rated as critical. This issue affects the function subprocess.run/subprocess.Popen of the file AWorld/aworld/virtual_environments/terminals/shell_tool.py. The manipulation leads to os command injection. The attack may be initiated remotely. The complexity of an attack is rather high. The exploitation is known to be difficult. The exploit has been disclosed to the public and may be used. This product does not use versioning. This is why information about affected and unaffected releases are unavailable.
python-markdownify (aka markdownify) before 0.14.1 allows large headline prefixes such as <h9999999> in addition to <h1> through <h6>. This causes memory consumption.
h11 accepts some malformed Chunked-Encoding bodies
h11 is a Python implementation of HTTP/1.1. Prior to version 0.16.0, a leniency in h11's parsing of line terminators in chunked-coding message bodies can lead to request smuggling vulnerabilities under certain conditions. This issue has been patched in version 0.16.0. Since exploitation requires the combination of buggy h11 with a buggy (reverse) proxy, fixing either component is sufficient to mitigate this issue.
The unsafe globals in Picklescan before 0.0.25 do not include ssl. Consequently, ssl.get_server_certificate can exfiltrate data via DNS after deserialization.
CVE-2025-24357 Malicious model remote code execution fix bypass with PyTorch < 2.6.0
## Description
https://github.com/vllm-project/vllm/security/advisories/GHSA-rh4j-5rhw-hr54 reported a vulnerability where loading a malicious model could result in code execution on the vllm host. The fix applied to specify `weights_only=True` to calls to `torch.load()` did not solve the problem prior to PyTorch 2.6.0.
PyTorch has issued a new CVE about this problem: https://github.com/advisories/GHSA-53q9-r3pm-6pq6
This means that versions of vLLM using PyTorch before 2.6.0 are vulnerable to this problem.
## Background Knowledge
When users install VLLM according to the official manual
 to aid remediation of the issue in the _ytdl-org/youtube-dl_ project.
### Vulnerability
_youtube-dl_ does not limit the extensions of downloaded files, which could lead to arbitrary filenames being created in the download folder (and path traversal on Windows).
### Impact
Since _youtube-dl_ also reads config from the working directory (and, on Windows, executables will be executed from the _youtube-dl_ directory by default) the vulnerability cou
Rasa Pro Missing Authentication For Voice Connector APIs
Rasa Pro is a framework for building scalable, dynamic conversational AI assistants that integrate large language models (LLMs). A vulnerability has been identified in Rasa Pro where voice connectors in Rasa Pro do not properly implement authentication even when a token is configured in the credentials.yml file. This could allow an attacker to submit voice data to the Rasa Pro assistant from an unauthenticated source. This issue has been patched for audiocodes, audiocodes_stream, and genesys connectors in versions 3.9.20, 3.10.19, 3.11.7 and 3.12.6.
Critical · CVSS 9.3EPSS 2.37%Public exploitFix available
PyTorch: `torch.load` with `weights_only=True` leads to remote code execution
PyTorch is a Python package that provides tensor computation with strong GPU acceleration and deep neural networks built on a tape-based autograd system. In version 2.5.1 and prior, a Remote Command Execution (RCE) vulnerability exists in PyTorch when loading a model using torch.load with weights_only=True. This issue has been patched in version 2.6.0.
Pycel through 1.0b30, when operating on an untrusted spreadsheet, allows code execution via a crafted formula in a cell, such as one beginning with the =IF(A1=200, eval("__import__('os').system( substring.
PyTorch LossCTC.cpp torch.nn.functional.ctc_loss denial of service
A vulnerability, which was classified as problematic, was found in PyTorch 2.6.0. Affected is the function torch.nn.functional.ctc_loss of the file aten/src/ATen/native/LossCTC.cpp. The manipulation leads to denial of service. An attack has to be approached locally. The exploit has been disclosed to the public and may be used. The real existence of this vulnerability is still doubted at the moment. The name of the patch is 46fc5d8e360127361211cb237d5f9eef0223e567. It is recommended to apply a patch to fix this issue. The security policy of the project warns to use unknown models which might establish malicious effects.
vLLM vulnerable to Denial of Service by abusing xgrammar cache
### Impact
This report is to highlight a vulnerability in XGrammar, a library used by the structured output feature in vLLM. The XGrammar advisory is here: https://github.com/mlc-ai/xgrammar/security/advisories/GHSA-389x-67px-mjg3
The [xgrammar](https://xgrammar.mlc.ai/docs/) library is the default backend used by vLLM to support structured output (a.k.a. guided decoding). Xgrammar provides a required, built-in cache for its compiled grammars stored in RAM. xgrammar is available by default through the OpenAI compatible API server with both the V0 and V1 engines.
A malicious user can send a stream of very short decoding requests with unique
Weblate VCS credentials included in URL parameters are potentially logged and saved into browser history as plaintext
Weblate is a web based localization tool. Prior to version 5.11, when creating a new component from an existing component that has a source code repository URL specified in settings, this URL is included in the client's URL parameters during the creation process. If, for example, the source code repository URL contains GitHub credentials, the confidential PAT and username are shown in plaintext and get saved into browser history. Moreover, if the request URL is logged, the credentials are written to logs in plaintext. If using Weblate official Docker image, nginx logs the URL and the token in plaintext. This issue is patched in version 5.11.
Jupyter Remote Desktop Proxy makes TigerVNC accessible via the network and not just via a UNIX socket as intended
Jupyter Remote Desktop Proxy allows you to run a Linux Desktop on a JupyterHub. jupyter-remote-desktop-proxy was meant to rely on UNIX sockets readable only by the current user since version 3.0.0, but when used with TigerVNC, the VNC server started by jupyter-remote-desktop-proxy were still accessible via the network. This vulnerability does not affect users having TurboVNC as the vncserver executable. This issue is fixed in 3.0.1.
Open Asset Import Library Assimp File MD3Loader.cpp ValidateSurfaceHeaderOffsets heap-based overflow
A vulnerability, which was classified as critical, was found in Open Asset Import Library Assimp 5.4.3. Affected is the function Assimp::MD3Importer::ValidateSurfaceHeaderOffsets of the file code/AssetLib/MD3/MD3Loader.cpp of the component File Handler. The manipulation leads to heap-based buffer overflow. The attack needs to be approached locally. The exploit has been disclosed to the public and may be used.
Denial of Service by abusing xgrammar unbounded cache in memory
XGrammar is an open-source library for efficient, flexible, and portable structured generation. Prior to 0.1.18, Xgrammar includes a cache for compiled grammars to increase performance with repeated use of the same grammar. This cache is held in memory. Since the cache is unbounded, a system making use of xgrammar can be abused to fill up a host's memory and case a denial of service. For example, sending many small requests to an LLM inference server with unique JSON schemas would eventually cause this denial of service to occur. This vulnerability is fixed in 0.1.18.