| CVE |
Vendors |
Products |
Updated |
CVSS v3.1 |
| Apache Kafka's code-generated protocol readers pre-size an array's collection from the element count declared by the sender. That count is checked only against the bytes remaining in the request buffer, which assumes about one byte per element; because most elements need several bytes, a request can declare far more elements than the buffer holds and still pass. For a keyed array the reader then allocates an ImplicitLinkedHashCollection of capacity (2 * count +1), so a request of a few tens of megabytes can force a heap allocation several times larger before any element is read. The pattern is emitted into every generated reader, so it affects all Kafka API types with an array field.
Impact:
Request parsing happens before any authorization check. On a listener that does not require authentication (for example PLAINTEXT), an unauthenticated attacker can send a few concurrent requests that each declare a very large=
array length and allocate many times their wire size in broker heap, crashing the broker with an OutOfMemoryError. The reporter demonstrated three concurrent ~90 MB ProduceRequest (v3) connections exhausting a 1 GB heap (~8x amplification). On listeners that require authentication, the attacker needs valid credentials but no specific ACL, since parsing precedes authorization..
Mitigation:
Upgrade to Apache Kafka 4.2.2, 4.3.2, or 4.4.0. Until then, lowering socket.request.max.bytes and restricting network access to broker listeners reduces, but does not remove, exposure.
Resolution:
From Apache Kafka 4.2.2, 4.3.2, and 4.4.0: generated readers cap the initial array pre-allocation at MessageUtil.MAX_PREALLOCATED_ARRAY_CAPACITY (1000) and grow it on demand, and reject a declared array length above MessageUtil.MAX_ARRAY_LENGTH (1,000,000). |
| The broker-side PushTelemetry API (KIP-714) validates only the compressed size of a client's metrics payload against telemetry.max.bytes (default 1 MB) before decompressing it. ClientTelemetryUtils.decompress() then decompresses the payload with no limit on the decompressed output size, growing a heap buffer until the compressed input is exhausted. A small, highly compressible payload (for example gzip-compressed zeros) can therefore expand to hundreds of megabytes of broker heap.
Impact:
An authenticated client can send a PushTelemetry request with a small compressed payload that decompresses to a very large size, consuming broker heap out of proportion to the request size. Repeated or concurrent requests can exhaust the broker heap and cause an OutOfMemoryError, degrading or crashing the broker; a client that targets multiple brokers can affect cluster availability. PushTelemetry is available to any authenticated client and is not gated by an ACL.
Mitigation:
Upgrade to Apache Kafka 4.2.2, 4.3.1, or 4.4.0. On affected versions there is no configuration that fully prevents the amplification, because telemetry.max.bytes bounds only the compressed size; lowering telemetry.max.bytes reduces the worst-case decompressed size proportionally and is a partial mitigation.
Resolution:
- From Apache Kafka 4.2.2, 4.3.1, and 4.4.0: ClientTelemetryUtils.decompress() enforces a maximum decompressed size equal to telemetry.max.bytes (default 1 MB). Decompression reads in bounded chunks and aborts with a TelemetryTooLargeException once the decompressed total exceeds the limit, so an over-large or highly compressible payload is rejected instead of being fully expanded in heap. The decompressed size is therefore bounded by default. |
| Allocation of resources without limits or throttling in Windows Netlogon allows an authorized attacker to deny service over a network. |
| Pillow is a Python imaging library. From 8.1.0 until 13.0.0, Pillow can allocate an APNG disposal background at the full canvas size before applying the decompression-bomb check. A crafted APNG with a small first frame, a very large declared canvas, and background disposal or first-frame OP_PREVIOUS reaches PngImageFile._seek while Image.open() positions frame zero. The full-canvas allocation occurs inside Image.open() without requiring pixel loading or frame iteration. This can exhaust memory and deny service. This issue is fixed in version 13.0.0. |
| AIOHTTP is an asynchronous HTTP client/server framework for asyncio and Python. Prior to 3.14.4, keepalive timeout handling is absent during the initial portion of a request. An attacker can keep an idle connection open indefinitely before sending any request data, consuming server connection capacity. This issue is fixed in version 3.14.4. |
| In Progress® Telerik® Document Processing Libraries prior to version 2026.3.1027, uncontrolled recursion, infinite loop, inefficient algorithmic complexity and resource allocation vulnerabilities in RadPdfProcessing and AI document summarization and question answering allow an unauthenticated attacker to terminate or hang the host process via a crafted PDF document or document text submitted for summarization or question answering. |
| vLLM is an inference and serving engine for large language models. From 0.19.0 until 0.32.0, authenticated callers can supply a request chat template when no server-side --chat-template is configured, or a malicious model repository can provide a template that is compiled and rendered by tokenizer.apply_chat_template() without a wall-clock timeout, output-size limit, or iteration bound. Rendering runs in a ThreadPoolExecutor with one worker by default, so a nonterminating render can block /v1/responses and other chat-template users after the request ends, while an output-expanding render can exhaust server memory. Harmony models are not affected because they do not use this Jinja path. This issue is fixed in version 0.32.0. |
| vLLM is an inference and serving engine for large language models. Prior to 0.32.0, the Phi-4 multimodal audio encoder in phi4mm_audio.py and phi4mm_utils.py uses eager attention by default, allowing a remote client to submit one legal long audio clip that causes MultiHeadedAttention.forward_attention to materialize and retain full attention matrices in self.attn. The scaled dot-product attention branch does not materialize these matrices. The resulting quadratic GPU-memory use can cause an out-of-memory condition and terminate EngineCore for all users. This issue is fixed in version 0.32.0. |
| vLLM is an inference and serving engine for large language models. From 0.16.0 until 0.32.0, the /v1/responses endpoint allows an authenticated caller to place an attacker-controlled template in chat_template_kwargs[chat_template] when no server template is configured. ChatParams.get_apply_chat_template_kwargs() preserves that value without enforcing trust_request_chat_template, so the request can bypass the intended trust policy and cause CPU or memory exhaustion through Jinja rendering. This issue is fixed in version 0.32.0. |
| Allocation of resources without limits in password-based private-key decryption (PbeUtilities.GenerateCipherParameters) in Legion of the Bouncy Castle Inc. bc-csharp before 2.7.0 allows an attacker who can supply an encrypted private key, such as a PKCS#8 EncryptedPrivateKeyInfo or "ENCRYPTED PRIVATE KEY" PEM file, to cause a denial of service through CPU exhaustion via an iteration count close to 2^31, because the count is taken from the unauthenticated algorithm parameters without an upper bound and the key derivation runs before the password or the data can be checked. PKCS#5 PBES1 and PBES2 (PBKDF2), the PKCS#12 PBE algorithms and CMS password recipients (CmsPbeKey) are affected. Loading PKCS#12 files with Pkcs12Store is covered by CVE-2026-63572, and a zero or negative count with the PKCS#12 algorithms by CVE-2026-63575. |
| A flaw was found in SSSD. In configurations where the autofs responder service is enabled, memory allocated during successful request processing is not released until the client connection terminates. A local attacker can exploit this vulnerability by maintaining an open connection and repeatedly submitting valid requests, leading to memory exhaustion and a Denial of Service (DoS). |
| In the Linux kernel, the following vulnerability has been resolved:
drm/amdgpu: keep the eviction fence until the last handle is closed
amdgpu_gem_object_close() detaches the user queue eviction fence from
the BO before it checks bo_va->ref_count, so closing any handle of a BO
removes the fence, even when the file still has another handle to the
BO and the bo_va with its GPU mappings stays. The BO can then be evicted
without suspending the user queues of the file, which keep accessing it
through mappings that are only marked for update.
Seen on navi31 with a SDMA user queue: a BO opened twice in the same
file (GEM_FLINK + GEM_OPEN) loses its eviction fence when the second
handle is closed (traced attach/detach of the fence):
attach bo=ffff8c78a6c84c00
GEM_CLOSE handle=7
detach bo=ffff8c78a6c84c00
while handle 6 and the mapping are still in use.
Only detach the fence when the last reference to the bo_va goes away,
matching amdgpu_gem_object_open(), which attaches it only when the
bo_va is created. With this the fence stays until the file is closed.
(cherry picked from commit 8387982f03d4026b7dc11b0b5e3118298eee4d6a) |
| Allocation of Resources Without Limits or Throttling vulnerability in the HTTP/2 request body reader of mtrudel bandit allows an unauthenticated remote attacker to exhaust server memory via a request body sent as many 1-byte DATA frames.
do_read_data/5 in lib/bandit/http2/stream.ex keeps every DATA frame payload as its own list element until the read returns. The :length option of read_body/2 counts body bytes, not frames, so a body of 1-byte frames costs a list cell and a binary term per byte. Interleaving the frames across up to 500 concurrent streams per connection keeps each stream consuming its frames as they arrive; in the reporter's test, 100 streams held 25 to 40 MiB of stream process memory for 5 MB sent. Only routes whose Plug reads the request body are affected, which includes any route behind Plug.Parsers.
This issue affects bandit: from 0.3.1 before 1.14.1. |
| Allocation of resources without limits or throttling in OpenSSH for Windows allows an unauthorized attacker to deny service over a network. |
| Allocation of resources without limits in PKCS#12 keystore loading (Pkcs12Store.Load) in Legion of the Bouncy Castle Inc. bc-csharp before 2.7.0 allows an attacker who can supply a PKCS#12 (PFX) file to cause a denial of service through CPU exhaustion via an iteration count close to 2^31 in the file's MacData or in the PBE parameters of an encrypted SafeContents or shrouded key bag, because the counts are taken from the file without an upper bound and the key derivation runs before the MAC or the password can be checked. A zero or negative count is covered by CVE-2026-63575. Pkcs12Utilities.ConvertToDefiniteLength is also affected. |
| Dell PowerProtect Data Domain, Feature Release versions 7.7.1.0 through 8.8, LTS2024 release versions 7.13.1.0 through 7.13.1.80, LTS2025 release versions 8.3.1.0 through 8.3.1.40, LTS2026 release versions 8.6.1.0 through 8.6.1.20, contain an Allocation of Resources Without Limits or Throttling vulnerability. An unauthenticated attacker with remote access could potentially exploit this vulnerability, leading to Denial of service. |
| Pillow is a Python imaging library. From 5.1.0 until 13.0.0, PdfParser.linearize_page_tree() in src/PIL/PdfParser.py treats a PDF /Pages graph as a tree without tracking duplicate indirect references. When Image.save(..., format="PDF", append=True) processes an attacker-controlled existing PDF whose /Kids arrays repeatedly reference the same child /Pages node, Pillow expands the shared subtree as distinct pages. A compact malformed PDF can therefore create exponentially many page references and consume disproportionate CPU before the append completes. This issue is fixed in version 13.0.0. |
| toml-node is a TOML parser for Node.js and the browser. Prior to 5.0.1, toml.parse() repeatedly slices and rejoins every path prefix in lib/compiler.js while processing long dotted keys, table headers, array-of-tables headers, and inline-table keys in assign(), deepRef(), and setNestedKey(). The resulting quadratic work can block the Node.js event loop for several seconds, and sufficiently long paths can terminate the process through memory exhaustion. The maxDepth option does not limit key-segment count and therefore does not mitigate this path. This issue is fixed in version 5.0.1. |
| pypdf is a free and open-source pure-python PDF library. Prior to 6.21.0, an attacker can craft a PDF containing arrays or dictionaries with many elements, causing parsing to consume large amounts of memory. This issue is fixed in version 6.21.0. |
| LibreSSL ocspcheck through 4.4.0 contains a memory exhaustion vulnerability that allows malicious OCSP responders to exhaust memory by sending unbounded HTTP responses. Attackers controlling the responder or intercepting plain HTTP AIA OCSP traffic can stream endless headers or body data to http_head_read() and http_body_read(), denying certificate status checks. |
