🚨 CVE-2026-114049
vLLM before 0.31.0 contains a cross-request mask leak in InputBatch.condense() that allows remote attackers to constrain other clients' outputs by leaving stale allowed_token_ids rows. Attackers can submit concurrent constrained generate requests, including via unauthenticated /invocations, so recycled batch slots apply their token allowlist to victims' unrestricted requests.

🎖@cveNotify

🚨 CVE-2026-114050
vLLM before 0.31.0 uses a client-supplied multimodal uuid verbatim as cache identity with no consistency check on cache hits, allowing engine crashes. Unauthenticated attackers can desynchronize the P0/P1 processor caches to trigger a fatal token-placeholder mismatch, killing the entire serving instance until restarted.

🎖@cveNotify

🚨 CVE-2026-114051
vLLM before 0.31.0 skips the min_tokens <= max_tokens check for /inference/v1/generate requests with a null max_tokens, so the EngineCore input thread later dies on an uncaught VLLMValidationError while /health still returns 200. Unauthenticated remote attackers can send one request with an oversized min_tokens to hang all subsequent generation requests until manual restart.

🎖@cveNotify

🚨 CVE-2026-114052
vLLM after 0.24.0 and before 0.31.0 contains a memory exhaustion vulnerability in MOSS-Audio model serving that caches a processor for every distinct request-controlled mm_processor_kwargs value in an unbounded dictionary. Unauthenticated attackers can send distinct or oversized mel_config values to the /tokenize endpoint, which bypasses --api-key, to exhaust host memory and crash the API process.

🎖@cveNotify

🚨 CVE-2026-114053
vLLM from 0.7.3 before 0.31.0 forwards the request-supplied mm_processor_kwargs.code_revision to Transformers AutoProcessor.from_pretrained, letting clients select which revision of a model's dynamic processor code is imported when trust_remote_code is enabled. An attacker controlling a revision in the served model repository can send a multimodal request to the unauthenticated /tokenize endpoint to execute arbitrary Python in the vLLM API process.

🎖@cveNotify

🚨 CVE-2026-114054
vLLM before 0.31.0 contains a cache aliasing vulnerability that allows authenticated attackers to collide shared cache keys by supplying caller-controlled values such as EXIF ImageID or cache_salt. Attackers can set a UUID EXIF ImageID to receive another request's cached processed media, or match cache_salt to a LoRA name to reuse first-block KV state.

🎖@cveNotify

🚨 CVE-2026-114055
vLLM from 0.7.3 before 0.31.0 contains an uncontrolled resource consumption vulnerability that allows remote clients to control multimodal processor parameters by supplying per-request mm_processor_kwargs. Attackers can set large image size, patch count, crop count, or frame count values in Isaac, InternVL, H2OVL, and other processors to exhaust memory and terminate the shared server.

🎖@cveNotify

🚨 CVE-2026-92822
A flaw was found in libmodulemd. When parsing module metadata, the xmd field is converted into nested GVariant values through recursive YAML mapping and sequence helpers without a recursion-depth or node-count limit. An attacker who can supply crafted modulemd YAML with deeply nested xmd content can cause stack exhaustion or excessive resource consumption, resulting in a denial of service.

🎖@cveNotify

🚨 CVE-2026-92823
A flaw was found in libmodulemd. In builds with rpmio support enabled, compressed modulemd YAML is decompressed through a custom read handler that does not enforce a cumulative decompressed-size limit before parsing. An attacker who can supply a highly compressible compressed YAML file can cause excessive CPU and memory consumption, resulting in a denial of service.

🎖@cveNotify

🚨 CVE-2026-92824
A flaw was found in libmodulemd. In the module stream v2 rpm-map emission path, an inner loop incorrectly uses the count of digest algorithms as its bound instead of the number of checksums in the current digest table. Crafted modulemd v2 metadata with multiple digest types can trigger an out-of-bounds read during YAML serialization, typically causing a crash and resulting in a denial of service.

🎖@cveNotify

🚨 CVE-2026-17557
A flaw was found in the `lm-evaluation-harness` component of Red Hat OpenShift AI. An authenticated user, able to submit an evaluation job, could exploit this by providing a malicious model URL in the `JobSpec`. This vulnerability allows for the disclosure of the tenant's model authentication secret to an attacker-controlled external listener. Additionally, it enables Server-Side Request Forgery (SSRF) attacks against internal network endpoints reachable from the evaluation job pod.

🎖@cveNotify

🚨 CVE-2026-109125
In NLnet Labs Unbound up to and including version 1.26.1, an observable cache-level side-channel can be exposed with 'unwanted-reply-threshold > 0' configurations (not a default configuration). If a response-like UDP datagram reaches the upstream source port currently used by a recursive query, but its TXID does not match any pending query, Unbound counts the packet as an unwanted reply. When the accumulated count reaches the configured threshold, Unbound clears the worker's RRset cache and message cache. This cache clear action does not destroy the current upstream UDP pending query. Therefore, an external observer can distinguish whether a candidate port matches the current upstream source port by checking whether an existing cached item has been cleared. This is a cache side-channel vulnerability that can be used to probe the upstream source port. Although the attack is not feasible on default configurations and the suggested value (10 million) makes it an ureliable attack on a public resolver, in scenarios where the attacker can keep a concurrent query window alive, the attacker can first filter the source port and then perform birthday-style guessing against the TXID, making cache poisoning a feasible attack path.

🎖@cveNotify

🚨 CVE-2026-114236
storage/pkg/archive.ApplyLayer and ApplyUncompressedLayer may, when processing corrupt or malicious inputs, leak a file descriptor; unbounded repeated calls can exhaust the process’ file descriptor
limit

🎖@cveNotify

🚨 CVE-2026-35501
When NLnet Labs Unbound 1.0.1 up to and including version 1.26.1 is configured with both a root trust anchor and an additional 'trust-anchor:' for an island-of-security zone whose parent is provably insecure under the root chain, a malicious actor with a forged NSEC reply with the SOA bit cleared can trick Unbound to treat the configured trust island as insecure. This happens because the DNSSEC validator in Unbound accepts and caches the NSEC insecure response bypassing the RRSIG validation that the validator was supposed to do. The acceptance of the NSEC record at the trust anchor apex was allowed due to and old fix for type DS and NSEC queries from clients when trust anchor for that name existed.

🎖@cveNotify

🚨 CVE-2026-50106
When NLnet Labs Unbound up to and including version 1.26.1 is configured with both a root trust anchor and an additional 'trust-anchor:' for an island-of-security zone whose parent is provably insecure under the root chain, a malicious actor who can deliver one forged NXDOMAIN response for a name inside the island, with the NSEC/SOA RRSIG 'SignerName' set to the insecure parent, tricks Unbound's validator to use the root trust anchor instead, and consecutively the insecure parent, thus causing that record to be treated as insecure and denying existence for records in the trust island.

🎖@cveNotify

🚨 CVE-2026-96405
NLnet Labs Unbound up to and including version 1.26.1 limits compression-pointer traversal for each individual domain-name operation but has no aggregate inbound decompression-work budget. Many chained compression pointers in a message results in a lot of CPU usage following them. A malicious actor with recursive-query access and control of an ordinarily delegated authority that serves specially crafted replies can exploit the vulnerability and degrade Unbound's general performance.

🎖@cveNotify

🚨 CVE-2026-97206
NLnet Labs Unbound, 1.22.0 up to and including version 1.26.1, contains a DNS-over-QUIC (DoQ) lifetime bug reachable by a remote DoQ client under multi-threaded operation. A client-triggered QUIC path migration can enter stream/connection cleanup while a mesh reply still retains a pointer to the DoQ stream. Cleanup frees the DoQ stream and when the pending mesh reply later completes it dereferences a stale pointer leading to a crash.

🎖@cveNotify

🚨 CVE-2026-105923
A malicious SVG can trigger an integer overflow which in turn triggers a null-pointer dereference in VectorImage QML Type.

🎖@cveNotify

🚨 CVE-2026-114392
Trudesk through 1.2.11 contains an authorization bypass vulnerability that allows authenticated users to read tickets outside their groups via GET /api/v1/tickets/:uid. Attackers can enumerate sequential ticket uids to read every ticket's subject, issue, comments, owner and assignee, because the apiTickets.single handler skips group membership checks.

🎖@cveNotify

🚨 CVE-2026-114394
Trudesk through 1.2.11 contains an authorization bypass vulnerability that allows authenticated users to post content attributed to other users by supplying user-controlled owner IDs. Attackers can set ownerId or owner in postComment, postInternalNote, and ticket create requests to forge administrator-authored comments, internal notes, and tickets on any ticket.

🎖@cveNotify

🚨 CVE-2026-114395
Trudesk through 1.2.11 contains an authorization bypass vulnerability in apiMessages.send that allows authenticated users to post messages as other users due to an incomplete CVE-2022-1931 fix. Attackers can supply arbitrary owner and cId values to POST /api/v1/messages/send to inject spoofed messages into conversations they do not belong to.

🎖@cveNotify