🚨 CVE-2026-100648
vllm before 0.29.0 fails to enforce VLLM_MAX_AUDIO_CLIP_FILESIZE_MB limit in multimodal chat audio decoding, allowing unauthenticated clients to bypass file size restrictions. Attackers can submit oversized audio files through chat endpoints to consume excessive memory and CPU resources during decoding.

🎖@cveNotify

🚨 CVE-2026-100649
vLLM before 0.29.0 contains a resource-limit bypass vulnerability in PyNvVideoCodec decoder allocation where sampler subclass shadowing allows independent counter increments. Unauthenticated attackers can select different sampler subclasses in video requests to exceed configured decoder limits and exhaust unaccounted GPU memory.

🎖@cveNotify

🚨 CVE-2026-100650
vLLM through 0.29.0 fetches and fully materializes remote or inline media before enforcing its documented media controls (the VLLM_MAX_AUDIO_CLIP_FILESIZE_MB compressed-audio size cap, default 25 MB, and the per-modality --limit-mm-per-prompt item limits). Across four ingress paths — the shared media-acquisition layer (HTTPConnection.get_bytes()/async_get_bytes()), the chat completions audio_url/base64 path, the batch speech runner, and the Rust frontend POST /tokenize route — the server reads the entire HTTP response body, base64-decodes the inline payload, or spawns one fetch/decode task per media part, and only then applies the limit (or, on some paths, never applies it). A remote attacker can therefore cause the API server or batch-runner process to allocate memory and consume outbound bandwidth proportional to an attacker-chosen body size or media item count before the request is rejected, resulting in pre-inference memory and bandwidth exhaustion (denial of service). The chat and batch surfaces require an API key when one is configured; the Rust frontend /tokenize route is unauthenticated by design. There is no code execution or data disclosure impact.

🎖@cveNotify

🚨 CVE-2026-100651
vLLM before 0.29.0 fails to enforce decoder prompt-length validation on the disaggregated serving endpoint /inference/v1/generate. When the request contains a 'features' (multimodal) payload, vllm/entrypoints/serve/disagg/serving.py builds a multimodal EngineInput directly from the caller-supplied token_ids, and GenerateRequest.token_ids (vllm/entrypoints/serve/disagg/protocol.py) is not checked against model_config.max_model_len. For multimodal processors that report skip_prompt_length_check=True (for example Nemotron Parse, Whisper, and FireRedLID), InputProcessor._validate_prompt_len() returns immediately for both encoder and decoder prompts, so an overlong prompt becomes an EngineCoreRequest and reaches the worker input-batch copy into a fixed max_model_len-wide NumPy row. A client able to reach the endpoint on an affected model configuration can therefore submit an overlong token_ids list to trigger a worker failure and denial of service. Fixed in 0.29.0.

🎖@cveNotify

🚨 CVE-2026-100652
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.

🎖@cveNotify

🚨 CVE-2026-100653
vLLM is an inference and serving engine for large language models. In versions from 0.22.1 through 0.28.0, the operator-supplied model revision pin (--revision / --code-revision) is not propagated to several Hugging Face artifact loads for the FunAudioChat and Tarsier2 architectures: the WhisperFeatureExtractor and speech_tokenizer PreTrainedTokenizerFast loads in vllm/model_executor/models/funaudiochat.py and the Qwen2VLConfig.from_pretrained call used by Tarsier2ProcessingInfo in vllm/model_executor/models/qwen2_vl.py. As a result, deployments pinned to a reviewed revision still resolve these behavior-affecting processor, tokenizer, and config artifacts from the repository's default revision, so a later change to the upstream default branch can alter audio preprocessing, speech tokenizer behavior, or Tarsier2 configuration without any change to the operator's configured pin. This is a supply-chain integrity and reproducibility failure for pinned deployments; it is residual to the earlier fix tracked as GHSA-3ww4-5jv9-j5gm / CVE-2026-47155 and does not constitute remote code execution or a trust_remote_code=False bypass. The issue is fixed in version 0.28.0.

🎖@cveNotify

🚨 CVE-2026-100654
vLLM before 0.29.0 accepts user-controlled stop_token_ids on the OpenAI-compatible POST /v1/completions and POST /v1/chat/completions endpoints but validates only that the values are integers, not that each token id is within the model vocabulary/logits range. When min_tokens > 0, the stop token ids are used as logits indices to suppress stop tokens, so an out-of-range id reaches a CUDA indexing operation (index_put_) and triggers a device-side assertion. An authenticated API user can send a single malformed completion request that returns 500 Internal Server Error and puts EngineCore into a fatal state, causing subsequent requests to fail until the service is restarted (denial of service).

🎖@cveNotify

🚨 CVE-2026-100656
Netty (io.netty:netty-codec-http) contains an unbounded per-connection queue growth flaw in HttpServerCodec. The codec tracks the HTTP method of each still-unanswered pipelined request; the first 32 entries are bit-packed into a single long, but every additional entry is appended to methodOverflowQueue, an ArrayDeque with no size limit and no rejection path. A remote, unauthenticated attacker who pipelines HTTP/1.1 requests on a single connection while withholding reads on their own end (preventing responses from being flushed) can grow this queue without bound, causing unbounded heap growth and denial of service. Affected versions are 4.2.0.Final through 4.2.17.Final and all releases up to and including 4.1.137.Final; the issue is fixed in 4.2.18.Final and 4.1.138.Final.

🎖@cveNotify

🚨 CVE-2026-100657
Netty's STOMP codec (io.netty:netty-codec-stomp) contains a ByteBuf leak in StompSubframeDecoder. Once a frame's declared content-length has been fully read, the decoder allocates a chunk buffer from the channel allocator and parks it in an instance field while waiting for the single NUL byte that terminates the frame. If that byte never arrives, the buffer is never released: the replay Signal thrown by skipNullCharacter extends Error rather than Exception, so the decoder's catch(Exception) release path does not run, and StompSubframeDecoder overrides neither handlerRemoved0 nor channelInactive, so the buffer also survives channel teardown. A remote peer can leak one allocator buffer per connection by sending a complete, well-formed frame body and withholding its terminating NUL byte; with the default pooled allocator the memory is never returned to the pool or reclaimed by garbage collection, so the leak accumulates for the lifetime of the process and can lead to memory exhaustion. This affects versions up to and including 4.1.137.Final and versions 4.2.0.Final through 4.2.17.Final; it is fixed in 4.1.138.Final and 4.2.18.Final.

🎖@cveNotify

🚨 CVE-2026-100658
Netty (io.netty:netty-codec-http) contains an unbounded per-connection queue in WebSocketServerExtensionHandler. The handler offers an entry to its per-channel validExtensions queue for every inbound HttpRequest, but polls an entry only when the application writes an HttpResponse, and the queue size is never bounded. A remote, unauthenticated peer can use HTTP/1.1 pipelining to send requests faster than the application produces responses — including plain non-upgrade HTTP requests to any path — causing the queue to grow without limit until the JVM exhausts heap memory and terminates with OutOfMemoryError. Because the affected handler is the base class of WebSocketServerCompressionHandler, any server that enables permessage-deflate is exposed on its plain HTTP port before any WebSocket upgrade completes and before any application-level authentication. Affected versions are 4.1.88.Final through 4.1.137.Final and 4.2.0.Final through 4.2.17.Final; the issue is fixed in 4.1.138.Final and 4.2.18.Final.

🎖@cveNotify

🚨 CVE-2026-100659
Netty's HTTP/3 codec (io.netty:netty-codec-http3) in versions 4.2.0.Final through 4.2.17.Final does not enforce the RFC 9114 requirement that the :authority pseudo-header field and a literal host header field, when both present, carry the same value. A remote unauthenticated peer can send a single HEADERS frame containing both fields with differing, attacker-controlled values; the request is accepted and delivered to the application with two conflicting authorities, allowing routing, virtual-host, and access-control decisions to be bypassed when different components in the request path consult different fields. This issue is fixed in 4.2.18.Final.

🎖@cveNotify

🚨 CVE-2026-100660
Netty's HTTP/3 codec (io.netty:netty-codec-http3) from 4.2.0.Final through 4.2.17.Final retains unbounded per-stream QPACK encoder state. QpackEncoder stores a queue and a dynamic-table index tracker for every encoded field section that references the QPACK dynamic table, keyed by the peer-controlled QUIC stream ID, and these entries are released only when the remote decoder sends a Section Acknowledgment or Stream Cancellation instruction — not when the HTTP/3 stream completes. There is no limit on the number of tracked streams, field sections, or retained bytes. A remote, unauthenticated HTTP/3 client can advertise a non-zero QPACK dynamic-table capacity, acknowledge the table insertion so the server reuses a dynamically indexed response header, and then omit all mandatory Section Acknowledgments while issuing sequential requests over a single QUIC connection, bypassing concurrent-stream limits and causing unbounded heap growth until the server exhausts memory (denial of service). Fixed in 4.2.18.Final.

🎖@cveNotify

🚨 CVE-2026-100661
Netty's HTTP/3 codec (io.netty:netty-codec-http3) versions 4.2.0.Final through 4.2.17.Final contain a denial-of-service vulnerability in the QPACK prefixed-integer decoder (QpackUtil.decodePrefixedInteger), which does not bound the number of continuation bytes it will process. A remote, unauthenticated peer can open a QPACK unidirectional stream (type 0x02 encoder or 0x03 decoder) and send a first byte with all prefix bits set (e.g. 0xFF for a 7-bit prefix or 0x3F for a 5-bit prefix) followed by an endless run of 0x80 continuation bytes. The decoder returns -1 ('need more bytes'), so callers never consume the input, the ByteToMessageDecoder cumulator grows without bound, and each decode() invocation re-scans the whole accumulated buffer, yielding O(N^2) CPU cost. The result is unbounded per-connection heap growth (OutOfMemoryError) and event-loop CPU starvation, reachable in every configuration. Fixed in 4.2.18.Final.

🎖@cveNotify

🚨 CVE-2026-100662
Netty's HTTP/3 codec (io.netty:netty-codec-http3) versions 4.2.0.Final through 4.2.17.Final contain an uncontrolled resource consumption vulnerability in the QPACK encoder-stream instruction decoder (QpackEncoderHandler, installed on the peer-initiated unidirectional QPACK encoder stream, type 0x02). The handler accepts an attacker-declared string-literal length of up to Integer.MAX_VALUE (~2 GiB) for the Name Length and Value Length fields of the "Insert With Literal Name" instruction (RFC 9204 §4.3.3), with no per-instruction or per-literal length cap and no cumulation-size limit; the existing HTTP/3 limits (maxHeaderListSize, maxUnknownFramePayloadLength, DEFAULT_MAX_FIELD_SECTION_SIZE) are not applied to this handler. A remote, unauthenticated peer with an established HTTP/3 connection to a default Netty HTTP/3 server can declare a very large literal length and then trickle fewer bytes than declared, causing the ByteToMessageDecoder MERGE cumulator to retain and grow the per-connection buffer, and ultimately triggering a large byte-array allocation. This leads to unbounded per-connection heap growth and OutOfMemoryError, resulting in denial of service. Fixed in 4.2.18.Final.

🎖@cveNotify

🚨 CVE-2026-100663
Netty's HTTP/3 codec (io.netty:netty-codec-http3) from 4.2.2.Final through 4.2.17.Final does not special-case HTTP/1 CONNECT authority-form request-targets when converting HTTP/1 messages to HTTP/3 in HttpConversionUtil.toHttp3Headers. The authority-form target (e.g., "CONNECT trusted.example:443") is parsed as a URI, so its host is emitted as :scheme, :path is set to "/", and the HTTP/1 Host header is used as :authority; if no Host header is present the CONNECT target is dropped. In a Netty-based HTTP/1-to-HTTP/3 proxy or gateway, a remote client can send a CONNECT request whose Host header names a different authority than the request-target, producing a malformed HTTP/3 CONNECT whose tunnel :authority is attacker-controlled. This can bypass tunnel allow-lists, egress policy, backend selection, or audit controls that validate the HTTP/1 CONNECT request-target before forwarding over HTTP/3. The issue is fixed in 4.2.18.Final.

🎖@cveNotify

🚨 CVE-2026-100664
Netty's HTTP/3 codec (io.netty:netty-codec-http3) versions 4.2.2.Final through 4.2.17.Final builds the HTTP/3 :authority pseudo-header from the HTTP/1 Host header before considering the authority of an absolute-form HTTP/1 request-target. In HttpConversionUtil.toHttp3Headers(HttpMessage, boolean) — reached via Http3FrameToHttpObjectCodec(false) — a non-empty Host header takes precedence over the request-target authority, contrary to the HTTP/1.1 rule that a server receiving an absolute-form request-target must ignore the Host header. In a Netty-based HTTP/1-to-HTTP/3 gateway, proxy, or protocol bridge, a remote client can send a request such as "GET https://trusted.example/admin HTTP/1.1" with "Host: attacker.example", causing components that validate, authorize, or route on the RFC-defined request-target authority to reach a different decision than the upstream HTTP/3 peer, which receives :authority derived from the conflicting Host header. This authority confusion can affect virtual-host routing, allow-list checks, backend selection, cache keys, and URL generation. The advisory reports integrity impact only (no code execution, memory corruption, or availability impact). Fixed in 4.2.18.Final.

🎖@cveNotify

🚨 CVE-2026-100665
Netty versions from 4.2.11.Final before 4.2.18.Final contain an incomplete hostname verification fix in the QUIC certificate verification path when using a plain X509TrustManager. The BoringSSLCertificateVerifyCallback discards the SSLEngine for plain trust managers, preventing endpoint identification from running even when HTTPS verification is configured. Attackers on the network path can present a certificate chain for the wrong hostname that the plain trust manager accepts, bypassing hostname authentication for QUIC clients.

🎖@cveNotify

🚨 CVE-2026-100666
Netty's HttpServerCodec (io.netty:netty-codec-http) in versions 4.2.0.Final through 4.2.16.Final and in versions up to and including 4.1.136.Final pairs each outbound response with an inbound request by calling pollMethod() once per response, including for 1xx informational responses. If a client pipelines an HTTP/1.1 GET carrying an Expect: 100-continue header followed by a HEAD request, the 100 Continue response consumes the queued GET method, so the subsequent 200 OK for the GET is paired with HEAD and its body is dropped, while the following 200 OK for the HEAD request is written with a body. This desynchronizes HTTP parsing on the connection: the GET entity is never delivered and the HEAD response body is interpreted as the GET body, resulting in response splitting and unsafe connection reuse. Fixed in 4.2.17.Final and 4.1.137.Final.

🎖@cveNotify

🚨 CVE-2026-100667
grav-plugin-login (the Grav CMS Login plugin) versions >= 3.8.7 and < 3.9.7 allow the two-factor authentication challenge to be bypassed for content gated by the authenticated() Twig function or the [authenticated] shortcode. On sites with 2FA enabled, Login::isAuthenticated() checked only the session flag indicating that the password step had succeeded, not the flag indicating that login had completed, so a session sitting at the 2FA code prompt was treated as fully authenticated. An attacker who knows a member's password but cannot answer that member's second factor can therefore read member-only content rendered by the no-argument authenticated() or group authenticated(null, 'group') forms and by [authenticated]; the inverse [guest] shortcode is likewise evaluated too early. Impact is limited to disclosure of that content: the attacker does not obtain a completed session, cannot access pages protected by an access: rule, and cannot act as the user. The authenticated('some.permission') form, which goes through UserObject::authorize(), is not affected. Fixed in grav-plugin-login 3.9.7.

🎖@cveNotify

🚨 CVE-2026-100668
Grav 2.0.0 through 2.0.24 contain a Twig content sandbox escape. The `array` filter (and its identical function form) is on the sandbox allowlist but is registered without the needs_is_sandboxed guard that print_r, vardump, json_encode, yaml_encode and string carry, and its implementation calls toArray() — or falls back to an (array) cast — without consulting the sandbox method allowlist. Because the `grav` Twig global is the raw Pimple-based dependency injection container, a user who can author Twig in page content can evaluate `grav|array` to read the container's private $values array, including the un-redacted Config service; a second array cast returns the entire configuration tree, disclosing plugin credentials, SMTP and OAuth secrets, Redis passwords, proxy URLs and the security.* subtree that the sandbox's redaction is meant to hide. Because the payload is stored in page content, the disclosed configuration is rendered to anonymous visitors. Grav 1.7 is not affected as it has no Twig content sandbox. Fixed in Grav 2.0.25.

🎖@cveNotify

🚨 CVE-2026-100669
Grav before 2.0.25 ships web server configuration samples whose access-control deny rules are matched case-sensitively. In webserver-configs/web.config (IIS), every deny rule (user_sensitive_folders, user_accounts, user_data, user_error_redirect, user_pages, system, vendor, ignore_folders) sets ignoreCase="false" on its URL Rewrite <match> element, overriding the IIS default of ignoreCase="true"; because these are rewrite matches rather than <requestFiltering> elements, there is no case-insensitive fallback. On IIS running over case-insensitive NTFS, an unauthenticated remote attacker can vary the case of a folder name or file extension (for example GET /user/CONFIG/system.YAML) so that no deny rule matches and the IIS static file handler resolves and returns the underlying file, disclosing sensitive data such as configuration secrets or account password hashes. Whether a bypassed file is actually returned depends on MIME registration: .json is served by default, while .yaml/.yml return HTTP 404.3 on a stock IIS unless a YAML MIME mapping has been added. The same class of gap exists in the bundled webserver-configs/lighttpd.conf, whose user/(config|env), directory, script-extension, root-file and dotfile rules lack the (?i) modifier, though it is lower risk because lighttpd typically runs on case-sensitive filesystems. Deployments served by Apache (.htaccess), nginx, Caddy, or the PHP built-in server are not affected. The issue is fixed in 2.0.25; because the .htaccess installer heal does not touch web.config or lighttpd.conf, operators must re-copy the corrected sample files after upgrading.

🎖@cveNotify