Vulnerabilities (CVE)

vLLM is an inference and serving engine for large language models. Prior to 0.30.0, Harmony tool continuations submitted through "POST /v1/responses" requests rebuild the next-turn engine input without preserving the cache_salt value, placing the continuation prefix in the global unsalted cache namespace even when the caller enabled salting. On deployments with prefix caching enabled, which is the default, an authenticated tenant who can reconstruct a victim's low-entropy post-tool history can submit the same continuation and use the cached_tokens_per_turn count to determine whether the prefix was previously processed, defeating the intended tenant isolation of salted prefix caching. This issue is fixed in version 0.30.0.Show less

vLLM is an inference and serving engine for large language models. Prior to 0.28.0, the default mirrored multimodal LRU cache can commit a media hash in the frontend sender cache during multimodal rendering and before engine admission, while the engine receiver cache never receives the payload if that request is rejected. A later request reusing the same media hash causes MultiModalProcessorSenderCache to send no payload and MultiModalReceiverCache to reach an assertion with the message "Expected a cached item," producing a shared-service availability failure. This issue is fixed in version 0.28.0.Show less

vLLM is an inference and serving engine for large language models. Prior to 0.30.0, the /inference/v1/generate endpoint in the disaggregated scale-out path accepts caller-supplied tensors in the features.kwargs_data field, cache identifiers in the features.mm_hashes field, ranges in the features.mm_placeholders field, and wire-selected multimodal field processors without rebinding them to the active model renderer contract. Forged grid geometry, field types, or non-positive placeholder lengths can terminate the shared EngineCore; when an attacker knows or can induce a victim's content hash, forged cache hashes can poison or retrieve cross-request encoder-cache state; and dropped sparse placeholder masks can alter replayed transport semantics. This issue is fixed in version 0.30.0.Show less

vLLM is an inference and serving engine for large language models. Prior to 0.30.0, flash late-interaction scoring at the /score and /rerank endpoints derives each worker's query_key value from the caller-controlled X-Request-Id header. A concurrent request that reuses a victim's identifier can overwrite the cached query embedding so the victim's documents are scored against the attacker's query, and shared use counters can also cause a late-interaction cache-miss error. This issue is fixed in version 0.30.0.Show less

vLLM is an inference and serving engine for large language models. Prior to 0.30.0, OpenAI-compatible request models accept a non-empty cache_salt value without enforcing the character and length restrictions required by the IPCCacheServerKey consumer in LMCache-MP. On deployments using the LMCache-MP connector, a salt that contains a forbidden character or exceeds the permitted length can raise an uncaught ValueError during scheduler cache lookup, causing EngineCore to terminate and denying service to all concurrent users. This issue is fixed in version 0.30.0.Show less

vLLM is an inference and serving engine for large language models. Prior to 0.30.0, structured-output request failures can escape request-scoped validation and reach the EngineCore fatal-error path. A per-request backend mismatch can re-raise a grammar compilation exception, padding produced by the ngram_gpu speculative-decoding mode can pass a negative token to guidance validation, and the Rust frontend can admit empty structured-output values that the Python frontend rejects, allowing ordinary constrained-generation requests to terminate the shared engine. This issue is fixed in version 0.30.0.Show less

vLLM is an inference and serving engine for large language models. From 0.24.0 until 0.30.0, the Qwen2VLVideoBackend and Qwen3VLVideoBackend classes accept request-level values for the media_io_kwargs.video.max_frames and media_io_kwargs.video.fps fields without enforcing server-side ceilings. An unauthenticated caller can submit these values to the /tokenize endpoint, causing the sampler to decode every frame selected from attacker-controlled video input, consume disproportionate frontend memory, and potentially terminate the API process before scheduling or admission control. The Rust frontend is not affected because it rejects the media_io_kwargs field. This issue is fixed in version 0.30.0.Show less

Confused deputy in WebAPKs in Google Chrome on on Android prior to 155.0.8059.39 allowed a remote attacker leveraging social engineering to bypass web origin policy via a crafted HTML page. (Chromium security severity: Low)Show less

IBM Langflow OSS 1.0.0 through 1.12.2 could allow a remote authenticated attacker to execute arbitrary OS commands due to improper neutralization of special elements used in an OS command ('Code Injection'), aka improper control of code generation.Show less

IBM Langflow OSS 1.0.0 through 1.12.2 could allow a remote authenticated attacker to execute arbitrary code due to improper neutralization of special elements used in an OS command ('Code Injection') related to improper input validation.Show less

Missing authorization in Google Lens in Google Chrome prior to 155.0.8059.39 allowed a remote attacker who had compromised the renderer process to obtain cross-origin data via a crafted HTML page. (Chromium security severity: Medium)Show less

Missing authorization in Autofill in Google Chrome prior to 155.0.8059.39 allowed a remote attacker leveraging social engineering to obtain sensitive information via a crafted HTML page. (Chromium security severity: Medium)Show less

Issue summary: A malicious remote peer may flood the local QUIC stack with NEW_CONNECTION_ID frames by avoiding a limit check on how many connection IDs the remote QUIC stack can use. Impact summary: The local QUIC stack sends a RETIRE_CONN_ID frame for every NEW_CONNECTION_ID frame it receives. The RETIRE_CONN_ID frame is dispatched via the Control Frame Queue (CFQ). If the remote peer also withholds ACKs, then it can force the local stack to allocate ~400MB (depending on ACK delay). CWE: CWE-770: Allocation of Resources Without Limits or Throttling Description: RFC 9000 sections 5.1.1 and 5.1.2 [1] describe the mechanism by which a remote peer can notify the local QUIC stack to change the destination connection ID (a.k.a. CID) the local stack uses to identify the connection at the remote peer. Each CID is associated with a sequence number. The sequence number is transmitted in NEW_CONNECTION_ID and RETIRE_CONNECTION_ID frames to identify the CID which is being either associated with a connection or retired. The remote peer sends a NEW_CONNECTION_ID frame to let the local stack know a new CID is being associated with an existing connection. The NEW_CONNECTION_ID frame carries the new CID, its sequence number, and the retire-prior-to number. The retire-prior-to identifies existing CIDs that are to be retired. The local QUIC stack must send a RETIRE_CONNECTION_ID for every destination CID whose sequence number is less than retire-prior-to. The CID becomes retired after the local stack receives an ACK for its RETIRE_CONNECTION_ID frame. Although the OpenSSL QUIC stack supports at most one destination CID for every connection, it can be tricked into processing more than one RETIRE_CONNECTION_ID frame per connection. The OpenSSL QUIC stack currently retires the destination CID as soon as it receives the NEW_CONNECTION_ID, while in fact the destination CID must be retired after an ACK for the RETIRE_CONNECTION_ID frame is received. Correcting the flawed logic also fixes the backlog growth. [1] https://datatracker.ietf.org/doc/html/rfc9000#name-issuing-connection-ids FIPS impact: no The FIPS module is not affected as the QUIC implementation is outside of the OpenSSL FIPS module boundary.Show less

Issue summary: The first concurrent use of the same X.509 certificate by several threads may cause its cached extension data to be freed while another thread is still using it. Impact summary: A remote, unauthenticated peer could crash a multi-threaded TLS client, or a multi-threaded TLS server that requests client certificates, if the first certificate chains built to the same trusted CA certificate are built by several connections at the same time. This is a use-after-free read, which is likely to crash the process, resulting in a Denial of Service. CWE: CWE-416: Use After Free Description: OpenSSL caches the decoded values of a certificate's X.509v3 extensions inside the X509 object the first time they are needed. In OpenSSL 4.0 this cache is built in two phases: the extension values are computed while holding a read lock on the certificate, and the results are then installed into the certificate under a write lock. Because a read lock does not exclude other readers, several threads can compute the cache for the same certificate at the same time. Each thread that subsequently acquires the write lock installs its own results and frees the values installed by the thread before it, even though that earlier thread has already marked the cache as complete and may have returned pointers into it to its caller. A caller still using those pointers then reads freed memory. Any certificate shared between threads is exposed the first time its extensions are decoded. In TLS the certificates at risk are the trusted CA certificates supplied for chain verification, by whatever means, since these are shared by every connection and their extensions are decoded and cached the first time a chain is built to them. Certificates sent by the peer are decoded separately for each connection and are not shared, so they are not affected. In a TLS client verifying server certificates, or a TLS server that requests and verifies client certificates, the use-after-free could only occur if the first chains built to the same trusted CA are built by several connections at the same time. FIPS impact: no The FIPS module is not affected as X.509 certificate handling is outside of the OpenSSL FIPS module boundary. OpenSSL 4.0 is vulnerable to this issue. OpenSSL 3.6, 3.5, 3.4, 3.0, 1.1.1 and 1.0.2 are not affected by this issue. OpenSSL 4.0 users should upgrade to OpenSSL 4.0.3. This issue was reported on 27 August 2026 by Tim Becker (Xint.io) and independently in a public report on 31 August 2026 by aydinmercan. The fix has been developed by Bob Beck. -- cut (non-publishing metadata for internal use) -- Reported by: Tim Becker (Xint.io), aydinmercan Fixed by: Bob BeckShow less