Recherche « Red Hat AI Inference Server »

Red Hat OpenShift AI 2.25 · Red Hat OpenShift AI 3.3 · Red Hat OpenShift AI 3.4 · Red Hat AI Inference Server

A flaw was found in Data Science Pipelines (DSP). An attacker with namespace editor privileges can bypass security hardening by submitting a malicious Argo Workflow through the V1 API path. This allows the API server to create pods with elevated privileges, acting as a 'confused deputy' on behalf of the attacker. Successful exploitation grants the attacker node-root access, enabling arbitrary code execution and full control over the underlying node.

xdgmime · Red Hat Enterprise Linux 10 · Red Hat Enterprise Linux 10.0 Extended Update Support · Red Hat Enterprise Linux 8

A flaw was found in xdgmime. A heap-based buffer overflow can be triggered in _xdg_mime_magic_parse_magic_line() in the xdgmimemagic.c file on little-endian systems when an attacker-controlled MIME magic file in a user-writable XDG data location (e.g., in the $XDG_DATA_HOME/mime/magic path) is parsed by an application performing MIME type detection (e.g., via g_content_type_guess()). When performing byte-swap, incorrect pointer arithmetic on the write side causes an out-of-bounds write of 2 bytes, resulting in an application crash or memory corruption.

huggingface text-generation-inference

text-generation-inference through 3.3.7 contains a server-side request forgery (SSRF) vulnerability in the OpenAI-compatible multimodal chat completions endpoint that allows unauthenticated network attackers to coerce the server into issuing arbitrary HTTP GET requests by supplying a crafted image_url value in chat message content. The fetch_image function in router/src/validation.rs performs no validation of private, loopback, link-local, or cloud metadata target addresses, and the reqwest HTTP client follows redirects by default, enabling attackers to bypass scheme checks via redirect chains to reach internal services and cloud instance-metadata endpoints for internal port scanning and credential theft.

GNOME GLib · Red Hat Enterprise Linux 10 · Red Hat Enterprise Linux 10.0 Extended Update Support · Red Hat Enterprise Linux 7 Extended Lifecycle Support

A flaw was found in GLib. The D-Bus client-side implementation of the DBUS_COOKIE_SHA1 SASL authentication mechanism does not validate the cookie_context parameter received from the server. A malicious D-Bus server can supply a cookie_context containing path traversal sequences, causing the client to read an arbitrary file and exfiltrate sensitive data by verifying guessed file contents against a generated hash.

GNOME GLib · Red Hat Enterprise Linux 10 · Red Hat Enterprise Linux 10.0 Extended Update Support · Red Hat Enterprise Linux 7 Extended Lifecycle Support

A flaw was found in GLib. A buffer over-read can occur in g_io_channel_read_line_backend() in the giochannel.c file when a custom line terminator with a length greater than one is set, causing memcmp to read past the GString buffer. This vulnerability can cause a minor information disclosure of 7 bytes or a denial of service when the buffer over-read crosses a page boundary.

GNOME GLib · Red Hat Enterprise Linux 10 · Red Hat Enterprise Linux 10.0 Extended Update Support · Red Hat Enterprise Linux 7 Extended Lifecycle Support

A flaw was found in GLib. A buffer over-read can occur in the g_regex_replace function when used with the `G_REGEX_RAW` compile flag and case-change replacement escapes because the string_append function processes matched substrings using UTF-8 functions that assume valid UTF-8 input, even when the string is treated as raw bytes. This vulnerability can cause a minor information disclosure of 1-5 bytes and a denial of service when the buffer over-read crosses a page boundary.

GNOME GLib · Red Hat Enterprise Linux 10 · Red Hat Enterprise Linux 10.0 Extended Update Support · Red Hat Enterprise Linux 7 Extended Lifecycle Support

A flaw was found in GLib. An out-of-bounds read of only 2 bytes can occur in the g_date_time_get_ymd function in the glib/gdatetime.c file when an invalid GDateTime object produced by the g_date_time_add_full function is processed. This flaw can corrupt the date output and potentially cause logic errors that may lead to a denial of service.

GNOME GLib · Red Hat Enterprise Linux 10 · Red Hat Enterprise Linux 10.0 Extended Update Support · Red Hat Enterprise Linux 7 Extended Lifecycle Support

A flaw was found in GLib. An off-by-one error can occur in the gvs_tuple_is_normal function in the glib/gvariant-serialiser.c file when doing an alignment padding check because the bounds check uses > instead of >=, causing an out-of-bounds read of only 1 byte. This issue can cause a minor information disclosure of 1 byte and a denial of service when the out-of-bounds read crosses a page boundary.

Red Hat Enterprise Linux 10 · Red Hat Enterprise Linux 10.0 Extended Update Support · Red Hat Enterprise Linux 9 · Red Hat Enterprise Linux 9.2 Update Services for SAP Solutions

A double free issue has been identified in libarchive's RAR5 reader. During parsing of a specially crafted RAR5 archive, the filtered_buf pointer may remain stale after being freed during unpacking state reinitialization. Subsequent processing of another archive entry can trigger a second free of the same memory region, resulting in a double-free condition. Successful exploitation may cause applications using the vulnerable libarchive API to terminate unexpectedly, leading to a denial of service.

FFmpeg · Red Hat Enterprise Linux AI 3.0 for RHEL 9 · Red Hat Enterprise Linux AI 3.2 for RHEL 9 · Red Hat Enterprise Linux AI 3.3 for RHEL 9

FFmpeg's RASC video decoder (decode_dlta in libavcodec/rasc.c) performs 32-bit reads and writes at the row cursor before the NEXT_LINE row-boundary check and validates the DLTA region in pixel rather than byte units, so a DLTA run on a PAL8 frame can access several bytes past the row allocation. A crafted media stream using the RASC FourCC, decoded by libavcodec, triggers a bitstream-controlled out-of-bounds heap write and adjacent out-of-bounds read, leading to memory corruption.

vllm-project vllm · Red Hat AI Inference Server 3.2 · Red Hat AI Inference Server 3.3 · Red Hat AI Inference Server 3.4

vLLM is an inference and serving engine for large language models (LLMs). Prior to 0.22.0, 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). This vulnerability is fixed in 0.22.0.

vLLM · Red Hat AI Inference Server · Red Hat Enterprise Linux AI (RHEL AI) 3 · Red Hat OpenShift AI (RHOAI)

vLLM versions >= 0.10.2 and < 0.13.0 are missing sparse tensor validation in multimodal embeddings processing. Because PyTorch disables sparse tensor invariant checks by default, an attacker can submit crafted embedding requests with malformed (negative or out-of-bounds) tensor indices, when the prompt-embeds feature is enabled, to trigger crashes or resource exhaustion (denial of service), with potential for out-of-bounds/write-what-where memory corruption. This continues CVE-2025-62164, whose prior fix only disabled the feature by default rather than addressing the root cause.

Red Hat Enterprise Linux AI 3.3 for RHEL 9 · Red Hat Enterprise Linux AI 3.4 for RHEL 9 · Red Hat Enterprise Linux AI 3.5 for RHEL 9 · Red Hat AI Inference Server 3.2

A remote code execution vulnerability was found in libaom, the reference AV1 codec implementation. Insufficient bounds validation in the AV1 encoder's SVC (Scalable Video Coding) layer ID control allows an attacker to supply crafted video frame pixels that overlap with internal encoder layer context structures. In fork-based video processing services, an attacker can use this to hijack the cyclic refresh map pointer, brute-force the process base address via a crash oracle, and redirect control flow to achieve arbitrary command execution. Exploitation requires the target service to use libaom with SVC encoding enabled and accept attacker-supplied video frames.

Red Hat Enterprise Linux AI 3.3 for RHEL 9 · Red Hat Enterprise Linux AI 3.4 for RHEL 9 · Red Hat Enterprise Linux AI 3.5 for RHEL 9 · Red Hat AI Inference Server 3.2

A heap-buffer-overflow read vulnerability was found in libaom, the reference AV1 codec implementation. A missing bounds check in the SVC (Scalable Video Coding) layer ID control function allows setting a spatial_layer_id exceeding the configured number of layers. This causes an out-of-bounds heap read of approximately 40,728 bytes when computing a layer context array index. An attacker who can influence SVC encoder parameters in a network-facing service could exploit this for information disclosure (heap content leak) or denial of service (segmentation fault from hitting unmapped memory).

Red Hat Enterprise Linux AI 3.3 for RHEL 9 · Red Hat Enterprise Linux AI 3.4 for RHEL 9 · Red Hat Enterprise Linux AI 3.5 for RHEL 9 · Red Hat AI Inference Server 3.2

An arbitrary address write vulnerability was found in libaom, the reference AV1 codec implementation. A missing bounds check in the SVC (Scalable Video Coding) layer ID control function allows an attacker to inject an arbitrary pointer into the cyclic refresh map field via crafted image pixel values. The encoder then writes approximately 1,200 bytes at the attacker-controlled address. This is fully deterministic and does not require a separate information leak. An attacker who can supply frames to a network-facing libaom encoder with SVC enabled could exploit this for denial of service or potential code execution.

Red Hat Enterprise Linux 10.0 Extended Update Support · Red Hat Enterprise Linux 7 Extended Lifecycle Support · Red Hat Enterprise Linux 8 · Red Hat Enterprise Linux 8.4 Advanced Mission Critical Update Support

A heap buffer overflow vulnerability was found in libaom, the reference AV1 codec implementation. A flaw in the AV1 encoder's Look-Ahead Processing (LAP) mode causes the first-pass stats ring buffer wrap-around guard to be bypassed when g_lag_in_frames is set to 1 or higher. This results in a 232-byte out-of-bounds write on every encoded frame after the second, corrupting adjacent heap objects. An attacker who can influence encoder configuration in a transcoding service or WebRTC session could exploit this to cause a denial of service (process crash) or potentially achieve code execution.

FFmpeg · Red Hat Enterprise Linux AI 3.5 for RHEL 9 · Red Hat AI Inference Server · Red Hat Enterprise Linux AI (RHEL AI) 3

An out-of-bounds write vulnerability in FFmpeg's libavcodec library, specifically in the MagicYUV decoder, allows denial-of-service and, in some cases, can be exploited for remote code execution. This vulnerability is associated with the file libavcodec/magicyuv.C. This issue affects FFmpeg before version 8.1.2.

Kludex starlette · Red Hat AI Inference Server 3.3 · Red Hat Exploit Intelligence · Red Hat Migration Toolkit for Applications 8

Starlette is a lightweight ASGI framework/toolkit. In versions 1.0.1 and earlier, StaticFiles on Windows is vulnerable to SSRF. An UNC path such as \\attacker.com\share can cause os.path.realpath to initiate an outbound SMB connection before the path is rejected, exposing the service account’s NTLMv2 credentials for offline cracking or relay even though the HTTP response is only a 404. The issue affects default follow_symlink=False deployments, including frameworks built on Starlette such as FastAPI; POSIX systems and follow_symlink=True are unaffected. The issue is fixed in 1.1.0.

vllm-project/vllm · Red Hat AI Inference Server · Red Hat Enterprise Linux AI (RHEL AI) 3 · Red Hat OpenShift AI (RHOAI)

vLLM versions 0.8.0 and later are vulnerable to an Out-of-Memory (OOM) Denial of Service (DoS) attack due to unbounded frame count processing in the `VideoMediaIO.load_base64()` method. When processing `video/jpeg` data URLs, the method splits the base64 data string on commas to extract individual JPEG frames without enforcing a frame count limit. An attacker can exploit this by crafting a single API request containing thousands of comma-separated base64-encoded JPEG frames in a data URL, causing the server to decode all frames into memory and crash due to excessive memory consumption. This vulnerability is reachable via the OpenAI-compatible chat completions API and does not require authentication.

aio-libs aiohttp · Red Hat Ansible Automation Platform 2.5 for RHEL 8 · Red Hat Ansible Automation Platform 2.5 for RHEL 9 · Red Hat Ansible Automation Platform 2.6 for RHEL 9

AIOHTTP is an asynchronous HTTP client/server framework for asyncio and Python. Prior to version 3.14.0, using ``CookieJar.load()`` with untrusted input may allow arbitrary code execution. Most applications using this function will be doing so with the user's own data, so this is unlikely to affect many applications. Version 3.14.0 patches the issue. If an application does allow attacker controlled files to be loaded, a workaround on older releases would be to sanitize the files before loading.

Red Hat Enterprise Linux 10 · Red Hat Enterprise Linux 10.0 Extended Update Support · Red Hat Enterprise Linux 7 Extended Lifecycle Support · Red Hat Enterprise Linux 8

A flaw was found in Poppler's Splash backend. A remote attacker could exploit this vulnerability by crafting a malicious PDF file that, when rendered, triggers an integer overflow in the `tilingPatternFill` function. This overflow leads to an undersized heap memory allocation, allowing a subsequent out-of-bounds write. Successful exploitation could result in arbitrary code execution, information disclosure, or denial of service within the context of the application processing the PDF.

jpadilla pyjwt · Red Hat Ansible Automation Platform 2.5 for RHEL 8 · Red Hat Ansible Automation Platform 2.5 for RHEL 9 · Red Hat Ansible Automation Platform 2.6 for RHEL 9

PyJWT is a JSON Web Token implementation in Python. Prior to 2.13.0, when the verifier is decoding JSON Web Tokens, while supporting both asymmetric and HMAC algorithms, the library does not validate use of JSON Web Keys in HMAC algorithm, allowing attacker to use the issuer public key as the secret key for HMAC algorithm. This vulnerability is fixed in 2.13.0.

open-webui · Openwebui open webui

Open WebUI is a self-hosted artificial intelligence platform designed to operate entirely offline. Prior to 0.9.0, Open WebUI supports model composition via base_model_id: a user-defined model (e.g., "Cheap Assistant") can reference an existing base model (e.g., "gpt-4-turbo-restricted") that provides the actual inference capability. When a user queries the composed model, the access control pipeline verifies the user has access to the composed model but never re-verifies access to the chained base model. Additionally, the model creation and import endpoints accept arbitrary base_model_id values without checking that the caller has access to that base model. Combined, this allows any user with the default model creation permission to create a model that chains to a restricted base model — and then invoke it, causing the server to dispatch the request to the restricted base model using the admin-configured API key. This vulnerability is fixed in 0.9.0.

huggingface diffusers · Red Hat OpenShift AI 3.4 · Red Hat AI Inference Server · Red Hat Enterprise Linux AI (RHEL AI) 3

Diffusers is the a library for pretrained diffusion models. Prior to 0.38.0, a trust_remote_code bypass in DiffusionPipeline.from_pretrained allows arbitrary remote code execution despite the user passing trust_remote_code=False (or omitting it, which is the default). The vulnerability has three variants, all sharing the same root cause — the trust_remote_code gate was implemented inside DiffusionPipeline.download() rather than at the actual dynamic-module load site, so any code path that bypassed or short-circuited download() also bypassed the security check. DiffusionPipeline.from_pretrained('repoA', custom_pipeline='attacker/repoB', trust_remote_code=False) — the gate evaluated against repoA's file list rather than repoB's, so repoB's pipeline.py was loaded and executed. DiffusionPipeline.from_pretrained('/local/snapshot', custom_pipeline='attacker/repoB', trust_remote_code=False) — the local-path branch never invoked download(), so the gate was never reached and remote code from repoB executed. DiffusionPipeline.from_pretrained('/local/snapshot', trust_remote_code=False) where the snapshot contains custom component files (e.g. unet/my_unet_model.py) referenced from model_index.json — same root cause; the local path skipped download() and custom component code executed. This vulnerability is fixed in 0.38.0.