CVE-2026-55574: Vllm Security Vulnerability (CVSS 8.7)

vLLM versions 0.22.0 through 0.23.0 fail to validate stop_token_ids against vocabulary bounds in Rust HTTP and gRPC frontends, allowing out-of-vocabulary token IDs to reach MinTokensLogitsProcessor. Attackers can submit requests with min_tokens greater than zero and out-of-vocabulary stop_token_ids to trigger CUDA tensor indexing failures that leave EngineCore in a fatal state requiring service restart.

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.

vLLM is an inference and serving engine for large language models (LLMs). The SSRF protection fix for CVE-2026-24779 add in 0.15.1 can be bypassed in the load_from_url_async method due to inconsistent URL parsing behavior between the validation layer and the actual HTTP client. The SSRF fix uses urllib3.util.parse_url() to validate and extract the hostname from user-provided URLs. However, load_from_url_async uses aiohttp for making the actual HTTP requests, and aiohttp internally uses the yarl library for URL parsing. This vulnerability in 0.17.0.

vLLM is an inference and serving engine for large language models (LLMs). Starting in version 0.10.1 and prior to version 0.14.0, vLLM loads Hugging Face `auto_map` dynamic modules during model resolution without gating on `trust_remote_code`, allowing attacker-controlled Python code in a model repo/path to execute at server startup. An attacker who can influence the model repo/path (local directory or remote Hugging Face repo) can achieve arbitrary code execution on the vLLM host during model load. This happens before any request handling and does not require API access. Version 0.14.0 fixes the issue.

The two built-in name-finder patterns exposed by opennlp.tools.namefind.RegexNameFinderFactory - DEFAULT_REGEX_NAME_FINDER.EMAIL and DEFAULT_REGEX_NAME_FINDER.URL - contain ambiguous nested quantifiers. An application that obtains these finders through RegexNameFinderFactory.getDefaultRegexNameFinders(...) and then applies them to untrusted text through RegexNameFinder.find(String[]) or RegexNameFinder.find(String) can be driven into super-linear backtracking or into unbounded matcher recursion by a small crafted input. For the EMAIL pattern, a long run of local-part characters that is never followed by an @ forces the matcher to re-scan to end-of-input from every starting offset. Cost grows quadratically with input length: an input of approximately 32 KB consumes several seconds of CPU in a single find() call and returns no match, and each doubling of the input multiplies the cost roughly four-fold. For the URL pattern, the query-string sub-expression nests a capturing repetition inside an outer repetition. The JDK matcher recurses once per query token, so an input of approximately 4 KB containing many &-separated tokens exhausts the thread stack and causes java.lang.StackOverflowError to propagate out of find(), terminating the calling thread. On a thread created with a smaller stack (for example -Xss512k, typical of server worker pools) approximately 1 KB is sufficient. In both cases an attacker who can supply text for analysis can convert a single request into seconds to minutes of pinned CPU, or into an abrupt thread death, denying service to the embedding application. No authentication, special configuration, or model file is required beyond the application having selected one of the two built-in finders. This issue affects Apache OpenNLP: from 2.0.0 through 2.5.11; from 3.0.0-M1 through 3.0.0-M5. Users are recommended to upgrade to version 2.5.12, or to 3.0.0-M6 for users tracking the 3.0.0 milestone line, which fix the issue.

Symfony is a PHP framework for web and console applications and a set of reusable PHP components. Prior to 5.4.52, 6.4.40, 7.4.12, and 8.0.12, Symfony\Component\Yaml\Parser::cleanup() used regular expressions with overlapping quantifiers for YAML directive, comment, and document marker cleanup, allowing crafted input to make parsing hang for an arbitrarily long time. This issue is fixed in versions 5.4.52, 6.4.40, 7.4.12, and 8.0.12.