Security Vulnerabilities
vLLM versions before 0.29.0 contain a denial-of-service vulnerability in the cache_salt parameter accepted on OpenAI-compatible and Anthropic API endpoints, which lacks maximum length validation and is processed on the single EngineCore scheduler thread. Unauthenticated attackers can send HTTP requests with multi-hundred-megabyte salt values that trigger expensive pickle serialization and SHA-256 hashing, stalling the scheduler thread and denying service to all concurrent requests.
CVSS Score
6.9
EPSS Score
0.003
Published
2026-09-26
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.
CVSS Score
6.9
EPSS Score
0.004
Published
2026-09-26
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.
CVSS Score
6.3
EPSS Score
0.003
Published
2026-09-26
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.
CVSS Score
7.1
EPSS Score
0.007
Published
2026-09-26
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.
CVSS Score
7.1
EPSS Score
0.003
Published
2026-09-26
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.
CVSS Score
8.2
EPSS Score
0.004
Published
2026-09-26
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.
CVSS Score
8.3
EPSS Score
0.003
Published
2026-09-26
In the Linux kernel, the following vulnerability has been resolved:
cgroup: Avoid iteration of dying tasks with zero refcount
The commit 260fbcb92bbea ("cgroup: Move dying_tasks cleanup from
cgroup_task_release() to cgroup_task_free()") extended the lifetime of
tasks on the dying_tasks list.
The iterators have provision to go through dying_tasks because of
dying threadgroup leaders or explicit CSS_TASK_ITER_WITH_DEAD, however,
it was expected that such tasks can obtain a new reference (that is
possible before cgroup_task_release()/put_task_struct_rcu_user()).
The tasks after cgroup_task_release() and before cgroup_task_free()
are subject to race when they may or may not have ->usage count > 0.
The race window is between css_task_iter_next() invocations
when css_set_lock is released and we may arrive at a new ->task_pos.
The iterator should not attempt to resurrect tasks whose ->usage count
dropped to zero. (When that happens, __put_task_struct_rcu_cb() is
already imminent and the returned task_struct would could be used
after free.)
As for the fix, we cannot simply check the signal->live count of a task
on the dying list because that won't distinguish regular zombies waiting
to be reaped from RCU remnant tasks that are going to be free'd.
Therefore add an extra check to rule out ->usage==0 tasks from any
iteration.
The repeat: loop in css_task_iter_advance() doesn't consider ->usage
count, so add a new loop to css_task_iter_next() to skip de-used tasks
on the dying_list.
Rough illustration of the possible race
R (reader of cgroup.procs) T (thread) L (group leader)
--------------------------------- -------------------------------- --------------------------------
L exits, signal->live > 0
cgroup_task_dead(L)
css_set_skip_task_iters() // skips only cset->tasks
list_add_tail(&L->cg_list, &cset->dying_tasks)
css_task_iter_next()
take css_set_lock
css_task_iter_advance()
leader && signal->live != 0
=> it->task_pos = &L->cg_list
release css_set_lock
T exits
--signal->live == 0
cgroup_task_dead(T) // css_set_lock
release_task(T)
cgroup_task_release(T)
release_task(L) // zap_leader
cgroup_task_release(L)
put_task_struct_rcu_user(L)
...RCU...
put_task_struct(L)
L->usage = 0
/* L still on dying_tasks */
...RCU...
__put_task_struct(L)
css_task_iter_next() // another iteration
take css_set_lock
it->task_pos = &L->cg_list
get_task_struct(L)
=> addition on 0
drop css_set_lock
cgroup_task_free(L)
css_set_skip_task_iters() // dying skip comes too late
free_task(L)
cgroup_procs_show()
task_pid_vnr(L)
CVSS Score
7.8
EPSS Score
0.001
Published
2026-09-26
Ghidra versions through 12.1.4 contain a stack-based out-of-bounds write vulnerability in the decompiler's leftshift128 function when processing negative shift amounts from p-code. Attackers can craft malicious binaries with specific instruction sequences that trigger the overflow when decompiled, corrupting memory and potentially achieving code execution.
CVSS Score
7.3
EPSS Score
0.002
Published
2026-09-26
Ghidra versions 9.2 through 12.1.4 contain a heap out-of-bounds read vulnerability in StringManager::getCodepoint when decoding multi-byte UTF-8, UTF-16, or UTF-32 characters without validating remaining buffer length. Attackers can craft malicious binaries containing strings or constant byte stores that end in multi-byte lead units to trigger out-of-bounds reads that crash the decompiler or leak adjacent heap memory into decompiled output.
CVSS Score
4.8
EPSS Score
0.001
Published
2026-09-26