VAITP

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.

When passing data to the b64decode(), standard_b64decode(), and urlsafe_b64decode() functions in the "base64" module the characters "+/" will always be accepted, regardless of the value of "altchars" parameter, typically used to establish an "alternative base64 alphabet" such as the URL safe alphabet. This behavior matches what is recommended in earlier base64 RFCs, but newer RFCs now recommend either dropping characters outside the specified base64 alphabet or raising an error. The old behavior has the possibility of causing data integrity issues. This behavior can only be insecure if your application uses an alternate base64 alphabet (without "+/"). If your application does not use the "altchars" parameter or the urlsafe_b64decode() function, then your application does not use an alternative base64 alphabet. The attached patches DOES NOT make the base64-decode behavior raise an error, as this would be a change in behavior and break existing programs. Instead, the patch deprecates the behavior which will be replaced with the newly recommended behavior in a future version of Python. Users are recommended to mitigate by verifying user-controlled inputs match the base64 alphabet they are expecting or verify that their application would not be affected if the b64decode() functions accepted "+" or "/" outside of altchars.

An undocumented and unsafe feature in the PLY (Python Lex-Yacc) library 3.11 allows Remote Code Execution (RCE) via the `picklefile` parameter in the `yacc()` function. This parameter accepts a `.pkl` file that is deserialized with `pickle.load()` without validation. Because `pickle` allows execution of embedded code via `__reduce__()`, an attacker can achieve code execution by passing a malicious pickle file. The parameter is not mentioned in official documentation or the GitHub repository, yet it is active in the PyPI version. This introduces a stealthy backdoor and persistence risk.

Using string formatting and exception handling, an attacker may bypass n8n's python-task-executor sandbox restrictions and run arbitrary unrestricted Python code in the underlying operating system. The vulnerability can be exploited via the Code block by an authenticated user with basic permissions and can lead to a full n8n instance takeover on instances operating under "Internal" execution mode. If the instance is operating under the "External" execution mode (ex. n8n's official Docker image) - arbitrary code execution occurs inside a Sidecar container and not the main node, which significantly reduces the vulnerability impact.

pyasn1 is a generic ASN.1 library for Python. Prior to 0.6.2, a Denial-of-Service issue has been found that leads to memory exhaustion from malformed RELATIVE-OID with excessive continuation octets. This vulnerability is fixed in 0.6.2.

Dask distributed is a distributed task scheduler for Dask. Prior to 2026.1.0, when Jupyter Lab, jupyter-server-proxy, and Dask distributed are all run together, it is possible to craft a URL which will result in code being executed by Jupyter due to a cross-side-scripting (XSS) bug in the Dask dashboard. It is possible for attackers to craft a phishing URL that assumes Jupyter Lab and Dask may be running on localhost and using default ports. If a user clicks on the malicious link it will open an error page in the Dask Dashboard via the Jupyter Lab proxy which will cause code to be executed by the default Jupyter Python kernel. This vulnerability is fixed in 2026.1.0.

Allocation of Resources Without Limits or Throttling in the HDF5 weight loading component in Google Keras 3.0.0 through 3.13.0 on all platforms allows a remote attacker to cause a Denial of Service (DoS) through memory exhaustion and a crash of the Python interpreter via a crafted .keras archive containing a valid model.weights.h5 file whose dataset declares an extremely large shape.

BlackSheep is an asynchronous web framework to build event based web applications with Python. Prior to 2.4.6, the HTTP Client implementation in BlackSheep is vulnerable to CRLF injection. Missing headers validation makes it possible for an attacker to modify the HTTP requests (e.g. insert a new header) or even create a new HTTP request. Exploitation requires developers to pass unsanitized user input directly into headers.The server part is not affected because BlackSheep delegates to an underlying ASGI server handling of response headers. This vulnerability is fixed in 2.4.6.

Deserialization of untrusted data in Azure Core shared client library for Python allows an authorized attacker to execute code over a network.

pypdf is a free and open-source pure-python PDF library. Prior to version 6.6.0, pypdf has possible long runtimes for missing /Root object with large /Size values. An attacker who uses this vulnerability can craft a PDF which leads to possibly long runtimes for actually invalid files. This can be achieved by omitting the /Root entry in the trailer, while using a rather large /Size value. Only the non-strict reading mode is affected. This issue has been patched in version 6.6.0.