vLLM: Dependency Confusion Vulnerability in vLLM Dockerfile
vLLM is an inference and serving engine for large language models (LLMs). Prior to 0.22.1, the vLLM Dockerfile is vulnerable to a dependency confusion attack through the flashinfer-jit-cache package. The package is installed from a custom index (flashinfer.ai/whl/) using --extra-index-url, but the package name was not registered on PyPI, and UV_INDEX_STRATEGY="unsafe-best-match" is set globally. An attacker who registers flashinfer-jit-cache on PyPI with version 0.6.11.post2 can execute arbitrary code as root during the Docker build and backdoor every resulting container image, enabling exfiltration of all user prompts, API credentials, and m
vLLM: OOM Denial of Service via Audio Decompression Bomb
vLLM is an inference and serving engine for large language models (LLMs). Prior to 0.23.1rc0, vLLM's /v1/audio/transcriptions endpoint limits compressed upload size but not decoded PCM output. A 25MB OPUS file expands to ~14.9GB of float32 PCM at decode time. This vulnerability is fixed in 0.23.1rc0.
vLLM is an inference and serving engine for large language models (LLMs). Prior to 0.23.1rc0, the fix for CVE-2026-22778, which introduced a sanitize_message helper that strips object-repr memory addresses from error messages before they reach the client, is incomplete: several response paths echo str(exc) directly to clients without calling sanitize_message. The unsanitized sites include the Anthropic API router in vllm/entrypoints/anthropic/api_router.py (the POST /v1/messages and POST /v1/messages/count_tokens handlers), the Server-Sent Events streaming converter in vllm/entrypoints/anthropic/serving.py, and the realtime speech-to-text Web
vLLM: temperature=NaN and temperature=Infinity bypass validation and propagate to GPU kernels
vLLM is an inference and serving engine for large language models (LLMs). Prior to 0.23.1rc0, ll temperature validation gates use comparison operators (<, >), which silently evaluate to False for NaN and for positive Infinity in Python's IEEE 754 float semantics. Both values pass every guard and propagate to GPU sampling kernels, where they produce undefined behavior or CUDA errors that can crash the inference worker. This vulnerability is fixed in 0.23.1rc0.
vLLM is an inference and serving engine for large language models (LLMs). From 0.3.0 until 0.22.0, a vulnerability in ASGI web servers and starlette's trust on those web servers enables an authentication bypass of the OpenAI API AuthenticationMiddleware. It allows to use the API without providing the configured VLLM_API_KEY or --api-key. This vulnerability is fixed in 0.22.0.
vLLM GGUF Kernels: int64_t to int truncation of tensor dimensions causes GPU buffer overflow
vLLM is an inference and serving engine for large language models (LLMs). From 0.5.5 until 0.23.1rc0, integer truncation of tensor dimensions in vLLM's GGUF dequantize kernels (csrc/quantization/gguf/gguf_kernel.cu) causes partial tensor processing. The output tensor is allocated at full size via torch::empty (uninitialized memory), but the dequantize CUDA kernel processes only a truncated number of elements. The unfilled portion of the output tensor retains whatever was previously in GPU memory. In multi-tenant inference deployments, this residual GPU memory may contain tensor data from other users' inference requests, constituting informati
WebOb: Location header normalization during redirect leads to open redirect
WebOb provides objects for HTTP requests and responses. Prior to 1.8.10, the normalization of the HTTP Location header during a redirect is vulnerable to an open redirect: WebOb joins the redirect target to the request URI using Python's urljoin, and since Python 3.10 the underlying urlsplit strips ASCII tab, carriage return, and newline characters before parsing, so a redirect target containing such characters can be reinterpreted as a protocol-relative URL whose authority is an attacker-controlled host. This bypasses the CVE-2024-42353 fix that escaped a leading double slash, allowing an attacker who influences the redirect location to send
picklescan - Remote Code Execution via idlelib.autocomplete.AutoComplete.get_entity
picklescan before 0.0.29 fails to detect malicious pickle files that exploit idlelib.autocomplete.AutoComplete.get_entity function in reduce methods. Attackers can embed undetected code in pickle files that executes arbitrary commands when loaded by victims using pickle.load().
picklescan - Arbitrary Code Execution via Undetected ensurepip._run_pip Function
picklescan before 0.0.30 (affected versions 0.0.26 and earlier) fails to detect the ensurepip._run_pip built-in function when scanning pickle files, allowing attackers to execute arbitrary code. Malicious pickle files embedding ensurepip._run_pip calls in __reduce__ methods bypass picklescan detection and achieve remote code execution upon pickle.load() invocation.
Picklescan - Arbitrary Code Execution via numpy.f2py.crackfortran._eval_length Gadget
Picklescan before 0.0.33 fails to detect the numpy.f2py.crackfortran._eval_length gadget in pickle __reduce__ methods, allowing arbitrary code execution. Attackers can craft malicious pickle files that execute arbitrary Python code when loaded by victims who trust Picklescan's safety validation.
UltraJSON: Malformed/Truncated UTF-8 Accepted and Silently Rewritten in ujson.dumps()
UltraJSON is a fast JSON encoder and decoder written in pure C with bindings for Python 3.7+. Prior to 5.13.0, ujson.dumps() (or ujson.dump() or ujson.encode()) have a reject_bytes=False option. When set, they may accept malformed or truncated UTF-8 byte sequences, silently rewriting them into different Unicode characters instead of rejecting them. This leads to input validation bypass and data integrity issues. This vulnerability is fixed in 5.13.0.
Critical · CVSS 9.5EPSS 3.03%Public exploitFix available
LiteLLM: Authentication Bypass via Host Header Injection
LiteLLM is a proxy server (AI Gateway) to call LLM APIs in OpenAI (or native) format. Prior to 1.84.0, a Host-header parsing flaw in the LiteLLM proxy could, under specific conditions, allow unauthenticated access to protected management routes. The auth layer derived the effective route from request.url.path in litellm/proxy/auth/auth_utils.py::get_request_route(), which Starlette reconstructs from the Host header. A crafted Host could therefore make the auth gate evaluate a different route from the one FastAPI dispatched. This vulnerability is fixed in 1.84.0.
Authlib OAuth 2.0 authorization endpoint open redirects to attacker-controlled redirect_uri on unsupported response_type
Authlib is a Python library which builds OAuth and OpenID Connect servers. Prior to 1.6.10 and 1.7.1, Authlib's OAuth 2.0 authorization endpoint can be turned into an unauthenticated open redirect when a request uses an unsupported response_type and supplies an attacker-controlled redirect_uri. The vulnerable behavior happens before client lookup and before any redirect URI validation. As a result, an attacker does not need a valid client registration, an authenticated user, or any prior state. A single request to the authorization endpoint is enough to obtain a 302 Location response to an arbitrary attacker-controlled URL. This vulnerability
pypdf: Possible infinite loop when processing threads/articles in writer
pypdf is a free and open-source pure-python PDF library. Prior to 6.13.1, an attacker who uses this vulnerability can craft a PDF which leads to an infinite loop. This requires merging a file with threads/articles into a writer. This vulnerability is fixed in 6.13.1.
pypdf: Inefficient decoding of FlateDecode PNG predictor streams
pypdf is a free and open-source pure-python PDF library. Prior to 6.12.2, an attacker who uses this vulnerability can craft a PDF which leads to long runtimes. This requires accessing a stream which uses the /FlateDecode filter with a PNG predictor. This vulnerability is fixed in 6.12.2.
pypdf: Possible large memory usage for form XObjects during text extraction
pypdf is a free and open-source pure-python PDF library. Prior to 6.12.2, an attacker who uses this vulnerability can craft a PDF which leads to large memory usage. This requires extracting the text of a page which contains a form XObject with self-references. This vulnerability is fixed in 6.12.2.
pypdf: Possible infinite loop when processing outlines/bookmarks in writer
pypdf is a free and open-source pure-python PDF library. Prior to 6.13.0, an attacker who uses this vulnerability can craft a PDF which leads to an infinite loop. This requires merging a file with outlines into a writer. This vulnerability is fixed in 6.13.0.
pypdf: Possible infinite loop when retrieving fonts for layout-mode text extraction
pypdf is a free and open-source pure-python PDF library. Prior to 6.13.0, an attacker who uses this vulnerability can craft a PDF which leads to an infinite loop. This requires extracting the text in layout mode. This vulnerability is fixed in 6.13.0.
Jupyter Server: Stored XSS in `NbconvertFileHandler` / `NbconvertPostHandler` via missing `sandbox` CSP
Jupyter Server is the backend for Jupyter web applications. Prior to 2.20, the nbconvert HTTP handlers in jupyter_server render user-authored notebook HTML under the Jupyter origin without a sandbox directive in their Content-Security-Policy. Combined with nbconvert.HTMLExporter's default non-sanitizing behavior, a notebook carrying an HTML payload in a display_data output triggers stored XSS with cookie access, full /api/* authority, and kernel RCE. This vulnerability is fixed in 2.20.
NLTK: URL-Encoded Path Traversal in nltk.data.load() Allows Arbitrary Local File Read
NLTK (Natural Language Toolkit) is a suite of open source Python modules, data sets, and tutorials supporting research and development in Natural Language Processing. Prior to 3.10.0-rc1, nltk.data.load() in NLTK is vulnerable to path traversal via URL-encoded path separators and traversal segments when using the nltk: URL scheme. The unsafe-path regex check is performed before url2pathname() decodes the %xx sequences (a classic decode-after-check / TOCTOU-style flaw), allowing an attacker to bypass the protection documented in NLTK's SECURITY.md and read arbitrary files from the filesystem. While literal traversal strings such as ../../../et