| CVE |
Vendors |
Products |
Updated |
CVSS v3.1 |
| vLLM versions before 0.28.0 fail to validate the lower bound of token IDs in the /v1/embeddings and /pooling endpoints, allowing unauthenticated attackers to crash the engine by submitting negative token IDs. A single request with a negative token ID triggers a CUDA device-side assertion that poisons the GPU context, causing all subsequent requests to fail until the process restarts. |
| Nodemailer versions >= 6.9.16 and < 9.1.0 mis-parse RFC 5322 comments in email addresses: in lib/addressparser, a comment closed immediately before a non-break character causes the tokenizer to concatenate the atoms surrounding the comment instead of treating the comment as folding whitespace that terminates the domain. A recipient address such as user@good-corp.com(x)evil.com is therefore read by Nodemailer as the single domain good-corp.comevil.com (registrable domain comevil.com, which an attacker can register) and used for both the SMTP envelope (RCPT TO) and the emitted To:/From: headers, while a conformant RFC 5322 parser terminates the domain at the comment and reads good-corp.com. An application that validates the recipient domain with a strict RFC 5322 parser (without inspecting parse defects) or a naive prefix/substring allow-list and then hands the raw address to Nodemailer can be induced to deliver mail to a domain the attacker controls. Fixed in 9.1.0. |
| Nodemailer (npm package `nodemailer`) versions 9.1.0 and earlier do not honor the `disableFileAccess` and `disableUrlAccess` sandbox options when message content is resolved through the public plugin API `MailMessage.resolveContent()` using the documented legacy three-argument signature `resolveContent(data, key, callback)`. Because `shared.resolveContent()` normalizes the missing `options` argument to an empty object, the message-level flags copied into `mail.data` by the MailMessage constructor are discarded, and `resolveContentValue()` skips both access-control checks, reaching `nmfetch(url)` or `fs.createReadStream(path)`. As a result, plugin or application code that resolves untrusted message content (html, text, attachment `path` or `href`) via this API can be induced to read arbitrary local files or issue outbound HTTP(S) requests (server-side request forgery), bypassing the sandbox the application enabled. The internal paths used by `transporter.sendMail()` (`resolveAll()`, `_convertDataImages()`, and the MIME streaming path) are not affected. Fixed in version 9.1.1. |
| vLLM through 0.29.0 contains a memory corruption vulnerability in the Triton _bincount_kernel where prompt token IDs index the penalty prompt-presence bitset without bounds checking against vocabulary size. Attackers can submit multimodal audio requests with tokens equal to vocabulary size, causing out-of-bounds writes that corrupt concurrent requests' sampler state and alter repetition penalty behavior. |
| rclone before 1.75.1 fails to confine names from server and third-party listing responses to the listed directory, allowing path traversal sequences in object names. Attackers can craft special names containing forward slashes and parent directory references to potentially write outside the destination root, though downstream protections in the local backend currently block actual file escape. |
| vLLM through 0.29.0 contains a denial of service vulnerability in the NIXL connector's prefix caching implementation that fails to properly validate block counts across multi-prompt completion requests in prefill/decode disaggregated deployments. Attackers can trigger an assertion failure in NixlBaseConnectorWorker._apply_prefix_caching by submitting completion requests with multiple prompts of varying lengths, causing the decode worker to terminate and become unavailable until restarted. |
| vLLM through 0.29.0 fails to validate the tp_size parameter in kv_transfer_params on OpenAI-compatible completion endpoints, allowing attackers to allocate unbounded memory. Attackers can supply arbitrary tp_size values in prefill/decode disaggregated deployments to exhaust memory and trigger kernel OOM-kill of the decode worker process. |
| vLLM through 0.29.0 fails to properly validate bad_words token indices against the model's generation output width in SamplingParams.update_from_tokenizer(). Attackers can supply out-of-bounds token indices that corrupt logits memory of concurrent requests, causing different in-flight HTTP requests to return incorrect tokens. |
| ImageMagick before 7.1.2-31 and before 6.9.13-56 contains a policy bypass in the PCD (and, per the upstream advisory, CUBE and HALD) coder: when a specific command line option is supplied, the decoder does not check a configured resource limit, which can result in extra memory allocation. A local user able to pass command line options to ImageMagick can therefore exceed the intended memory policy limit, causing a limited availability impact. The issue is fixed in 7.1.2-31 and 6.9.13-56. |
| ImageMagick before 7.1.2-31 contains a policy bypass vulnerability in the UHDR encoder that fails to perform policy checks during buffer allocation for image pixels. Attackers can bypass resource policies by processing specially crafted UHDR images, potentially causing denial of service through excessive memory allocation. |
| uutils coreutils versions before 0.10.0 apply setuid or setgid mode to install destinations before finalizing ownership changes, allowing privileged users to leave setuid executables owned by the privileged invoker when ownership changes fail. Attackers can execute leftover setuid files with elevated privileges when ownership change operations fail on capability-restricted systems. |
| ImageMagick before 7.1.2-31 and before 6.9.13-56 contains a NULL pointer dereference in the PNM coder. When the coder reaches a memory (resource) limit at a specific point during processing, the failed allocation is not handled and a NULL pointer is dereferenced, which can lead to a denial of service (application crash) when processing a specially crafted or sufficiently large PNM image. |
| braces through 3.0.3 contains a stack overflow vulnerability in the recursive AST walkers that lack depth guards. Attackers can supply deeply nested brace patterns under the character limit to exhaust the call stack and terminate the Node.js process with an uncaught RangeError. |
| Expat through 2.8.4 fails to validate low surrogates following high surrogates in UTF-16 input, allowing malformed UTF-16 sequences to be accepted. Attackers can craft UTF-16 encoded XML with lone high surrogates that consume following code units, hiding markup characters from the parser and enabling XML injection attacks. |
| vLLM through 0.29.0 contains a denial of service vulnerability in P2P KV offloading when OffloadingConnector is configured with TieringOffloadingSpec and a peer-to-peer secondary tier. Attackers can supply arbitrary remote host and port values in kv_transfer_params to create unreachable peer sessions that retain ZeroMQ sockets until the context quota is exhausted, causing an uncaught ZMQError that crashes EngineCore and stops all inference. |
| vLLM Mooncake connector through 0.29.0 fails to properly manage GPU KV cache block ownership when concurrent child requests share a single transfer ID in prefill/decode disaggregated deployments. Attackers can trigger GPU memory exhaustion by submitting completion requests with multiple prompts, causing orphaned KV cache blocks to accumulate until process restart and eventually preventing legitimate requests from executing. |
| vLLM before 0.29.0 validates allowed_token_ids against tokenizer length instead of model output logits width in SamplingParams._validate_allowed_token_ids(). Attackers can supply token IDs above the output vocabulary that pass validation, causing LogitBiasState to corrupt GPU logits state and allow concurrent requests to sample tokens outside their allowlists. |
| vLLM versions through 0.29.0 contain a denial of service vulnerability in the NIXL connector's metadata handling for prefill/decode disaggregated deployments. Attackers can send requests with incomplete kv_transfer_params dictionary entries to trigger an uncaught KeyError in EngineCore scheduling, causing the decode engine to terminate and making all routed requests fail until manual restart. |
| joi (npm package `joi`, hapi.js) versions >=17.2.0 <17.13.7 and >=18.0.0 <18.2.6 are vulnerable to regular expression denial of service in the `Joi.string().isoDate()` validation rule. One of the regular expressions the rule applies to the input is unanchored, so a valid ISO date followed by a long run of fractional-second digits causes the regex engine to restart its search from every position in the string, yielding time proportional to the square of the input length (about 1.4 s for 64 KB of digits and about 22 s for 256 KB). A remote attacker who can supply a string to an isoDate validation can stall the application with a single request. Fixed in 17.13.7 and 18.2.6; as a workaround, cap the length of the string before it reaches joi. |
| vLLM through 0.29.0 fails to properly clean up decode-side metadata for rejected inference requests in prefill/decode disaggregated deployments. Remote attackers can submit requests with max_tokens=0 to exhaust decode-worker memory without bound until the worker restarts. |
