| CVE |
Vendors |
Products |
Updated |
CVSS v3.1 |
| SGLang through 0.5.21 contains an unbounded memory allocation vulnerability in preprocess_video in qwen_vl.py that decodes frames at full source resolution before applying max_pixels limits. Remote attackers can submit high-resolution videos to /v1/chat/completions on Qwen-VL deployments, and concurrent requests exhaust worker memory causing denial of service. |
| mcp-go through 1.2.1 contains a denial of service vulnerability in StreamableHTTPServer.ServeHTTP that allows remote unauthenticated attackers to exhaust memory by sending oversized POST bodies. Attackers can send arbitrarily large or many concurrent POST requests, read fully via io.ReadAll before validation, to degrade or OOM-kill the server process. |
| Allocation of resources without limits or throttling in Active Directory Federation Services (AD FS) allows an unauthorized attacker to deny service over a network. |
| A vulnerability was found in AUTOMATIC1111 stable-diffusion-webui up to 1.10.1. This impacts the function read_metadata_from_safetensors of the file modules/sd_models.py of the component Metadata Extraction. The manipulation of the argument header length field results in allocation of resources. It is possible to launch the attack remotely. The exploit has been made public and could be used. The project was informed of the problem early through an issue report but has not responded yet. |
| Allocation of resources without limits or throttling in Active Directory Federation Services (AD FS) allows an unauthorized attacker to deny service over a network. |
| StreamReadConstraints.maxNameLength (default 50,000 characters) is intended to bound how much memory a single JSON Object property name can consume while being parsed, independently of the much larger maxStringLength (default 100,000,000 characters) that bounds ordinary string values. For parsers built over character input — Reader, String or char[] sources, all served by ReaderBasedJsonParser — that bound is not applied while the name is being accumulated. The code path instead re-uses the string-value length check, which is roughly 2000x larger by default.
In the character-input slow path, ReaderBasedJsonParser._parseName2() accumulates the name into the TextBuffer and, when a segment fills, calls TextBuffer.finishCurrentSegment(), which validates only maxStringLength. The correct check, StreamReadConstraints.validateNameLength(int), is reached afterwards in CharsToNameCanonicalizer.findSymbol(), once the entire name has been scanned to its closing quote. By that point the full name — up to maxStringLength characters — has already been buffered in memory.
A remote attacker who can have a single sufficiently long JSON Object property name parsed through this path can force allocation on the order of 100,000,000 characters (approximately 200 MB, with segment bookkeeping pushing real peak usage higher) before any StreamConstraintsException is thrown, instead of the approximately 50,000-character (about 100 KB) bound the API promises. Repeating this across concurrent requests can exhaust the heap and crash the process. No authentication or user interaction is required. Because StreamReadConstraints.maxDocumentLength is disabled by default (-1), nothing else bounds the token first, so the issue is exploitable under default configuration.
Byte-input parsers are not affected: UTF8StreamJsonParser, UTF8DataInputJsonParser and the non-blocking (async) parsers re-validate maxNameLength on every growth of the name-decode buffer via ParserBase._growNameDecodeBuffer(), rejecting an oversized name after allocating only tens of kilobytes. Applications are exposed when untrusted JSON is parsed from a String, Reader or char[] — for example an HTTP request body already read into a String — rather than from raw UTF-8 bytes.
This issue affects com.fasterxml.jackson.core:jackson-core from 2.16.0 and tools.jackson.core:jackson-core from 3.0.0. Releases before 2.16.0 are not affected, because StreamReadConstraints.maxNameLength was first introduced in jackson-core 2.16.0. |
| RabbitMQ is a messaging and streaming broker. From 4.0.0 until 4.0.24, 4.1.15, 4.2.10, and 4.3.4, deps/amqp10_common/src/amqp10_binary_parser.erl materializes each element of AMQP 1.0 arrays whose shared constructor has zero width, while limiting only the number of elements in each array and not the aggregate arrays in a frame. After completing SASL authentication, a low-privilege client can send an ordinary frame containing many individually permitted arrays, causing a small frame to expand into a very large reader-process heap before later frame validation or authorization. Because deps/rabbit/src/rabbit_amqp_reader.erl clears the pre-authentication max_heap_size guard after login, repeated or concurrent frames can trigger the node-wide memory alarm, block publishers, or exhaust broker memory. This issue is fixed in versions 4.0.24, 4.1.15, 4.2.10, and 4.3.4. |
| Mattermost versions 11.10.x <= 11.10.0, 11.9.x <= 11.9.1, 11.8.x <= 11.8.5, 11.7.x <= 11.7.10 fail to cap the number and size of mention keys stored in a user's notification properties, and to bound the quadratic multibyte-keyword matching performed for each post, which allows an authenticated user to degrade post-processing latency and consume excessive CPU on behalf of other members of any channel they share via storing an oversized notify_props.mention_keys list on their own profile.. Mattermost Advisory ID: MMSA-2026-00773 |
| Nezha from 2.0.13 before 2.3.15 contains a resource exhaustion vulnerability that allows authenticated users to open unbounded concurrent connections to the /api/v1/ws/transfer WebSocket endpoint. Attackers can hold thousands of subscriptions created by ServerTransferClass.Subscribe without limits, exhausting file descriptors, memory, and goroutines and blocking agent gRPC reconnections. |
| Nezha 1.0.0 before 2.3.15 contains a memory exhaustion vulnerability in the dashboard file manager WebSocket endpoint because fmStream never calls SetReadLimit, leaving message size unbounded. Authenticated users with file manager scopes can send one oversized binary message to /api/v1/ws/file/:id, triggering an OOM kill that drops all agent monitoring. |
| The Model Context Protocol Python SDK is the official Python SDK for Model Context Protocol servers and clients. In mcp versions before 1.29.1 and from 2.0.0a1 through 2.0.1, the Streamable HTTP session manager, SSE message endpoint, and OAuth authorization-server routes buffer complete HTTP request bodies without enforcing a maximum size. An unauthenticated remote client able to reach an affected endpoint can send a large declared or streamed request body that consumes multiple times the body size in process memory and can repeatedly exhaust memory or terminate the server. Stdio-only servers and deployments with an effective upstream request-size limit are not affected, and applications that use StreamableHTTPServerTransport directly remain responsible for enforcing their own limit. The standard SDK server applications and session managers are fixed in versions 1.29.1 and 2.1.0. |
| Pipecat is an open-source Python framework for building real-time voice and multimodal conversational agents. From 0.0.98 until 1.6.0, the pipecat-ai package's built-in development runner exposes unauthenticated endpoints including /start, /daily, and /daily-dialin-webhook that can create bot sessions and Daily rooms. The runner binds to localhost by default and is intended only for local development, so exploitation requires an operator to expose it to untrusted traffic through --host 0.0.0.0, a published port, or a tunnel without access controls. An unauthenticated caller can consume local resources and third-party quotas and, depending on configuration, receive a room URL and join token. Version 1.6.0 adds a prominent development-runner banner and documentation that clarifies the unauthenticated boundary and directs production deployments to an authenticated dispatcher or managed runtime, while the endpoint behavior remains by design. This issue is fixed in version 1.6.0. |
| In Eclipse Vert.x versions [4.0.0, 4.5.34] or [5.0.0,5.0.12] or [5.1.0,5.1.9] or [5.2.0, 5.2.0], the Vert.x Redis client RESP parser pre-allocates arrays to a length declared by the server without enforcing an upper bound. A malicious or compromised Redis server, or an attacker in a man-in-the-middle position, can send a crafted RESP response declaring an extremely large array, push, or attribute length, causing the client to attempt a massive heap allocation. The resulting OutOfMemoryError crashes the Vert.x event loop thread, denying service to all connections and verticles sharing that loop.
The impacted component is io.vertx:vertx-mysql-client. |
| An Allocation of Resources Without Limits or Throttling vulnerability in the l2cpd process of Juniper Networks Junos OS allows an adjacent unauthenticated attacker to cause a denial of service and l2cpd restart via FNP-enabled Layer 2 interfaces. The issue is triggered when the FNP feature is enabled on a VPLS routing instance with VLAN-VPLS encapsulation enabled.
This issue affects all platforms except the PTX Series and SRX Series. |
| Text::CSV_XS versions before 1.65 for Perl allow memory exhaustion by allocating one entry per index of an RFC 7111 selector in fragment.
fragment expands a row= or col= selector into an array with one element per selected index before it reads any data, and it does not clamp an explicit end column in a cell= selector to the row. Only the syntax and order of the ranges are checked, so row=1-400000000 on a two-line file allocates 400 million elements.
An application that passes a request-supplied fragment to fragment or csv can be made to exhaust memory, which ends the process with an Out of memory error that eval cannot catch. |
| Text::CSV_XS versions before 1.65 for Perl allow quadratic CPU use by copying the unparsed remainder per record of CR-only input in getline.
With eol unset, the default, getline reads with Perl's default line separator, so input with CR-only line endings arrives as one chunk. After each record the parser copies the whole unparsed remainder into a new buffer, which the next call parses and copies again, so the work grows with the square of the input size.
A reader with default options that is sent CSV with classic Mac line endings spends seconds of CPU time on a megabyte of input. |
| In Eclipse Vert.x versions [4.0.0, 4.5.34] or [5.0.0,5.0.12] or [5.1.0,5.1.9] or [5.2.0, 5.2.0], the MySQL message decoder aggregates inbound multi-packet payloads into a composite direct-memory buffer without any cumulative size limit. A malicious, compromised, or man-in-the-middle MySQL server can send a continuous stream of packets with the maximum payload length (0xFFFFFF), preventing the end-of-message condition from being reached, causing unbounded direct-memory allocation on the client and leading to an out-of-memory crash. This occurs before authentication completes.
The impacted component is io.vertx:vertx-mysql-client. |
| In `api-gateway` versions till `1.7.3`, The default rate-limit key resolver, `ClientIpKeyResolver`, derives the per-client bucket key from the leftmost element of the `X-Forwarded-For` request header. That element is supplied by the originating HTTP client and is not validated against any trusted-proxy list. A single attacker can therefore obtain an unlimited number of independent rate-limit buckets — one per unique header value — and completely bypass the gateway's request-rate limiting on every route that uses the shipped default configuration. |
| undici is an HTTP client library for Node.js. In versions 6.5.0 through 6.29.0, 7.0.0 through 7.30.0, and 8.0.0 through 8.12.1, its EventSource client can be driven to exhaust process memory by a malicious or compromised Server-Sent Events peer. By sending millions of tiny completed data fields with no blank event separator, or a single unfinished line split across many chunks, the parser retains a growing amount of per-field and per-chunk state that the byte-based maxEventSize and maxResponseSize limits do not bound, and because maxResponseSize is enforced on the compressed body before decompression, a small compressed response can expand into enough retained memory to terminate the process. An application only needs to maintain an EventSource connection to an untrusted peer, which controls only the response, and availability is the only impact. The issue is fixed in undici 6.29.1, 7.30.1, and 8.12.2, which bound the retained parser state independently of the byte limits, and users should upgrade to one of those versions or later. There is no workaround. |
| vLLM is an inference and serving engine for large language models. Prior to 0.30.0, the Rust frontend's track_http_metrics middleware records the raw HTTP method token as a Prometheus label for requests reaching registered routes. An unauthenticated attacker can send unique arbitrary method tokens to unguarded routes such as /tokenize, causing Prometheus's Family::get_or_create function to permanently create counter and histogram label sets. Those label sets increase process memory usage and enlarge the /metrics response until the service or monitoring path is exhausted. This issue is fixed in version 0.30.0. |
