GHSA-8397-8qf6-f2gm CVE-2026-115913

vLLM: Phi-4 multimodal audio encoder materializes and retains full attention matrices: a single max-length audio OOM-crashes the vLLM engine

Medium PyPI

Severity: 6.5 / 10; Medium · CVSS 3.1
Exploit chance (EPSS): 0.33%; Chance of exploitation in the next 30 days. Higher than 26.12% of scored vulnerabilities.
Known exploited: No; Not in CISA's catalog
Weakness: CWE-770 Allocation of Resources Without Limits or Throttling
Fix: 0.32.0; Upgrade vllm to this version or later

Description

Summary

The Phi-4 multimodal audio encoder (phi4mm_audio.py / phi4mm_utils.py) runs its conformer attention in eager mode by default: the official HF config of microsoft/Phi-4-multimodal-instruct does not set use_pt_scaled_dot_product_attention, so the default False applies (phi4mm_audio.py:166, :876; phi4mm_utils.py:1669). The eager path fully materializes the attention probability matrix of shape [n_chunks, 16, 500, 500] per layer — and then stores it on the module: self.attn = attn (phi4mm_utils.py:1802, a torch.jit.Attribute declared at :1693), keeping a live reference for every one of the 24 encoder layers.

The only gates on audio length are the server-side limits VLLM_MAX_AUDIO_CLIP_FILESIZE_MB=25 and VLLM_MAX_AUDIO_DECODE_DURATION_S=600 (envs.py:83-84) — i.e. a 10-minute clip is a legal input by default. A 600 s clip becomes ~60,000 mel frames, which the encoder unfolds into 500-frame chunks (phi4mm_audio.py:996, max_seq_len = 500), giving n_chunks ≈ 120. Each layer therefore materializes a 120×16×500×500 matrix (~1–2 GiB depending on dtype), and the module attribute keeps all completed layers' matrices alive, so GPU memory climbs monotonically within a single request until an allocation fails.

Verified end-to-end on the official vllm/vllm-openai:v0.31.0 image: one 19.2 MB silent 600 s WAV → worker torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 2.29 GiB ... 501.75 MiB free ... 22.09 GiB allocated by PyTorch inside get_audio_features (phi4mm_audio.py:1279 → :1332 → phi4mm_utils.py:1386) → EngineCore encountered a fatal error → API EngineDeadError → whole server down; /health refuses connections. Total time from request to dead service: ~6 seconds.

A secondary effect: because self.attn persists on the module after the request ends, the matrices of the last audio request stay resident indefinitely (observed +50 MiB pinned after a single 5 s control request); they are only overwritten by the next audio request, so the resident footprint equals the largest audio processed since startup.

Details

1. Root cause. MultiHeadedAttention.forward_attention in eager mode computes attn = masked_softmax(scores, mask) and unconditionally executes self.attn = attn (phi4mm_utils.py:1795-1802). The attribute is a leftover inference/debug hook (used by torch.jit.Attribute(None, Tensor | None) at :1693); in serving it serves no purpose but pins the full [batch·chunks, heads, T, T] tensor per layer past both the layer's forward and the request's lifetime. The SDP attention branch (use_pt_scaled_dot_product_attention=True) never materializes this matrix — but the HF config does not enable it and the code default is False.

2. Why one request suffices. The conformer chunks long audio to max_seq_len = 500 frames (phi4mm_audio.py:996). Duration is bounded only by VLLM_MAX_AUDIO_DECODE_DURATION_S=600, so n_chunks reaches ~120 for a maximal clip. With 16 heads and fp32 scores, one layer's attention tensor is ~1.8 GiB (the observed failing allocation was 2.29 GiB, consistent with the scores+bias materialization). The encoder has num_blocks=24, attention_heads=16, attention_dim=1024, time_reduction=8 (HF config). Weights plus a handful of retained per-layer matrices exhaust the ~2.3 GiB headroom that --gpu-memory-utilization 0.9 leaves on a 24 GB card — mid-forward, on the first request.

4. Deployment reality. --trust-remote-code is the documented, standard way to serve this model (the HF model card instructs it), so the finding applies to default deployments. The crash is deterministic, not load-dependent: scheduler stats at death show a single running request and near-zero KV cache usage.

Impact

Any unauthenticated client that can reach the chat/completions API of a default deployment serving Phi-4-multimodal-instruct (or any model reusing this audio tower) can kill the entire engine with one request carrying a perfectly legal 10-minute audio clip (silent noise suffices — no crafted content needed). Full denial of service, requiring a process restart. On GPUs with more headroom (e.g. 80 GB at default utilization) a single request may survive, but the eager path still materializes GiB-scale per-layer matrices and retains the last request's tensors indefinitely, so a few sequential max-length requests — or normal traffic following one — achieve the same result.

Relation to prior art: the only public Phi-4-mm advisory we found, GHSA-vc6m-hm49-g9qg, concerns a CPU-side quadratic-cost path — different subsystem, different mechanism. GitHub search for "phi4mm audio OOM/memory" (31 results, 2026-10-08) shows no report of this attention-retention issue; main (2026-10-08) is unchanged.

Suggested remediation (any one breaks the chain; 1+2 recommended):

Remove self.attn = attn from the serving path (or gate it behind an explicit debug/output_attentions flag) so attention matrices are freed when the layer returns.
Default use_pt_scaled_dot_product_attention to True (flash/SDPA kernels never materialize the [B, H, T, T] matrix); keep eager only as an opt-in.
Additionally bound the encoder input (frames or duration) for this model so the eager path cannot reach GiB-scale single-layer allocations even if enabled.

This advisory was fixed in MR60614.

Affected versions

PyPI

Package	Vulnerable	Fixed in
vllm	`< 0.32.0`	`0.32.0`
vllm	`< 0.32.0`	`0.32.0`

How to fix

Move to the fixed version or later with your package manager, update the lockfile, and run your tests. If the package only arrives through another dependency, upgrade that dependency first. If none of its releases pulls a fixed version yet, force the version with your package manager's override.

vllm PyPI → `0.32.0`

pip

pip install --upgrade "vllm>=0.32.0"

poetry

poetry add "vllm>=0.32.0"

uv

uv add "vllm>=0.32.0"

Severity breakdown

CVSS 3.1 6.5

CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H

Attack vector: Network
Attack complexity: Low
Privileges required: Low
User interaction: None
Scope: Unchanged
Confidentiality: None
Integrity: None
Availability: High

Weakness

CWE-770: Allocation of Resources Without Limits or Throttling

The product allocates a reusable resource or group of resources on behalf of an actor without imposing any intended restrictions on the size or number of resources that can be allocated.

How to prevent it

Requirements: Clearly specify the minimum and maximum expectations for capabilities, and dictate which behaviors are acceptable when resource allocation reaches limits.
Architecture and Design: Limit the amount of resources that are accessible to unprivileged users. Set per-user limits for resources. Allow the system administrator to define these limits. Be careful to avoid CWE-410.
Architecture and Design: Design throttling mechanisms into the system architecture. The best protection is to limit the amount of resources that an unauthorized user can cause to be expended. A strong authentication and access control model will help prevent such attacks from occurring in the first place, and it will help the administrator to identify who is committing the abuse. The login application should be protected against DoS attacks as much as possible. Limiting the database access, perhaps by caching result sets, can help minimize the resources expended. To further limit the potential for a DoS attack, consider tracking the rate of requests received from users and blocking requests that exceed a defined rate threshold.
Implementation: Assume all input is malicious. Use an "accept known good" input validation strategy, i.e., use a list of acceptable inputs that strictly conform to specifications. Reject any input that does not strictly conform to specifications, or transform it into something that does. When performing input validation, consider all potentially relevant properties, including length, type of input, the full range of acceptable values, missing or extra inputs, syntax, consistency across related fields, and conformance to business rules. As an example of business rule logic, "boat" may be syntactically valid because it only contains alphanumeric characters, but it is not valid if the input is only expected to contain colors such as "red" or "blue." Do not rely exclusively on looking for malicious or malformed inputs. This is likely to miss at least one undesirable input, especially if the code's environment changes. This can give attackers enough room to bypass the intended validation. However, denylists can be useful for detecting potential attacks or determining which inputs are so malformed that they should be rejected outright.
Architecture and Design: For any security checks that are performed on the client side, ensure that these checks are duplicated on the server side, in order to avoid CWE-602. Attackers can bypass the client-side checks by modifying values after the checks have been performed, or by changing the client to remove the client-side checks entirely. Then, these modified values would be submitted to the server.
Architecture and Design: Mitigation of resource exhaustion attacks requires that the target system either: recognizes the attack and denies that user further access for a given amount of time, typically by using increasing time delays uniformly throttles all requests in order to make it more difficult to consume resources more quickly than they can again be freed. The first of these solutions is an issue in itself though, since it may allow attackers to prevent the use of the system by a particular valid user. If the attacker impersonates the valid user, they may be able to prevent the user from accessing the server in question. The second solution can be difficult to effectively institute -- and even when properly done, it does not provide a full solution. It simply requires more resources on the part of the attacker.
Architecture and Design: Ensure that protocols have specific limits of scale placed on them.
Architecture and Design, Implementation: If the program must fail, ensure that it fails gracefully (fails closed). There may be a temptation to simply let the program fail poorly in cases such as low memory conditions, but an attacker may be able to assert control before the software has fully exited. Alternately, an uncontrolled failure could cause cascading problems with other downstream components; for example, the program could send a signal to a downstream process so the process immediately knows that a problem has occurred and has a better chance of recovery. Ensure that all failures in resource allocation place the system into a safe posture.
Operation, Architecture and Design: Use quotas or other resource-limiting settings provided by the operating system or environment. For example, when managing system resources in POSIX, setrlimit() can be used to set limits for certain types of resources, and getrlimit() can determine how many resources are available. However, these functions are not available on all operating systems. When the current levels get close to the maximum that is defined for the application (see CWE-770), then limit the allocation of further resources to privileged users; alternately, begin releasing resources for less-privileged users. While this mitigation may protect the system from attack, it will not necessarily stop attackers from adversely impacting other users. Ensure that the application performs the appropriate error checks and error handling in case resources become unavailable (CWE-703).

References

Advisories

Fixes and patches

Packages

github.com/vllm-project/vllm

Timeline

Oct 28, 2026 Published
Oct 28, 2026 Last updated by the source
Oct 31, 2026 Latest EPSS score

Advisory	Severity	Why it's related
GHSA-x567-49g5-p4cf vLLM: Responses API bypasses trust_request_chat_template for chat_template_kwargs	Medium	Also affects vllm
GHSA-vvpg-rrww-wpmh vLLM: Permissive default CORS origins allow cross-origin requests to the vLLM HTTP server	Medium	Also affects vllm
GHSA-q97c-p368-75c4 vLLM: Out-of-vocabulary prompt token id passes input validation and triggers a CUDA device-side assert (engine crash / DoS)	Medium	Also affects vllm
GHSA-h3rc-j82p-2c3h vLLM: Untrusted EPS inputs can trigger denial of service or Ghostscript RCE	High	Also affects vllm
GHSA-gq5q-89c7-vhvx vLLM DeepSeek-OCR image processing vulnerable to denial of service	Medium	Also affects vllm
GHSA-6757-hggm-hp2w vLLM: Public multimodal SHM handle forgery reaches worker-side `pickle.loads()` with insecure serialization disabled	High	Also affects vllm
GHSA-542m-9234-8gvf vLLM: DoS via unbounded chat_template Jinja2 rendering exhausts the renderer thread pool	Medium	Also affects vllm
PYSEC-2026-4353 vLLM from 0.20.0 before 0.30.0 contains a denial of service vulnerability that allows remote attackers to crash the engine by sending inconsistent multimodal…	High	Also affects vllm

Credits: C1ves, jperezdealgaba

Is this in your code?

MergeHarbor checks every repository's dependencies against this database, ranks what matters by exploitability and reachability, routes it to the owning team and opens the fix pull request.

Check your repos with MergeHarbor