Skip to content

vLLM denial of service via prompt embeds on M-RoPE models

High severity GitHub Reviewed Published Jul 2, 2026 in vllm-project/vllm • Updated Jul 20, 2026

Package

pip vllm (pip)

Affected versions

>= 0.12.0, < 0.24.0

Patched versions

0.24.0

Description

Summary

Short summary of the problem. Make the impact and severity as clear as possible. For example: An unsafe deserialization vulnerability allows any unauthenticated user to execute arbitrary code on the server.

Sending a pure prompt embeds payload in a /v1/completions request with a model using M-RoPE causes the EngineCore to fail an assertion and fatally crash, shutting down the entire server application.

Any remote user who is authorized to make a /v1/completions endpoint can trivially make such a request and induce a crash.

Details

Give all details on the vulnerability. Pointing to the incriminated source code is very helpful for the maintainer.

In commit 56669c1, a simple assert intended to be a type-narrowing assert was added to the _init_mrope_positions method in GPUModelRunner (the offending line on main at the time of writing: https://github.com/vllm-project/vllm/blob/2d481f8a946ee0521872af0f098674a8ee01ce4a/vllm/v1/worker/gpu_model_runner.py#L1588-L1607).

assert req_state.prompt_token_ids is not None, (
            "M-RoPE requires prompt_token_ids to be available."
        )

This type narrowing assert is to prevent mypy errors later in the function because None is not a valid type for mrope_model.get_mrope_input_positions. Unfortunately, this assertion is not always true. /v1/completions requests that specify prompt=None and prompt_embeds=<not none> will indeed create a CachedRequestState where prompt_token_ids is None. This triggers the assertion, which in turn crashes the EngineCore and the Server application.

(EngineCore pid=351) ERROR 06-11 00:48:03 [core.py:1167]   File "/usr/local/lib/python3.12/dist-packages/vllm/v1/worker/gpu_model_runner.py", line 3997, in execute_model
(EngineCore pid=351) ERROR 06-11 00:48:03 [core.py:1167]     deferred_state_corrections_fn = self._update_states(scheduler_output)
(EngineCore pid=351) ERROR 06-11 00:48:03 [core.py:1167]                                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore pid=351) ERROR 06-11 00:48:03 [core.py:1167]   File "/usr/local/lib/python3.12/dist-packages/vllm/v1/worker/gpu_model_runner.py", line 1239, in _update_states
(EngineCore pid=351) ERROR 06-11 00:48:03 [core.py:1167]     self._init_mrope_positions(req_state)
(EngineCore pid=351) ERROR 06-11 00:48:03 [core.py:1167]   File "/usr/local/lib/python3.12/dist-packages/vllm/v1/worker/gpu_model_runner.py", line 1582, in _init_mrope_positions
(EngineCore pid=351) ERROR 06-11 00:48:03 [core.py:1167]     assert req_state.prompt_token_ids is not None, (
(EngineCore pid=351) ERROR 06-11 00:48:03 [core.py:1167]            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore pid=351) ERROR 06-11 00:48:03 [core.py:1167] AssertionError: M-RoPE requires prompt_token_ids to be available.
(APIServer pid=1) ERROR 06-11 00:48:03 [async_llm.py:704] AsyncLLM output_handler failed.
(APIServer pid=1) ERROR 06-11 00:48:03 [async_llm.py:704] Traceback (most recent call last):
(APIServer pid=1) ERROR 06-11 00:48:03 [async_llm.py:704]   File "/usr/local/lib/python3.12/dist-packages/vllm/v1/engine/async_llm.py", line 660, in output_handler
(APIServer pid=1) ERROR 06-11 00:48:03 [async_llm.py:704]     outputs = await engine_core.get_output_async()
(APIServer pid=1) ERROR 06-11 00:48:03 [async_llm.py:704]               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(APIServer pid=1) ERROR 06-11 00:48:03 [async_llm.py:704]   File "/usr/local/lib/python3.12/dist-packages/vllm/v1/engine/core_client.py", line 1030, in get_output_async
(APIServer pid=1) ERROR 06-11 00:48:03 [async_llm.py:704]     raise self._format_exception(outputs) from None
(APIServer pid=1) ERROR 06-11 00:48:03 [async_llm.py:704] vllm.v1.engine.exceptions.EngineDeadError: EngineCore encountered an issue. See stack trace (above) for the root cause.

All requests using the /v1/chat/completions endpoints will have text/prompt_token_ids parts (corresponding to the chat template), and prompt_embeds parts are handled as mm_features. This method (rightly) filters out those prompt_embeds content parts as they are treated as text positions.

A sufficient solution to type narrowing here without raising a fatal assertion is to instead replace the assertion with a using dummy token ids:

def _init_mrope_positions(self, req_state: CachedRequestState):
    model = self.get_model()
    assert supports_mrope(model), "M-RoPE support is not implemented."
    mrope_model = cast(SupportsMRoPE, model)

    # Filter out prompt_embeds modality (text-only position info)
    mrope_features = [
        f for f in req_state.mm_features if f.modality != "prompt_embeds"
    ]
    
    # Handle both token_ids and embeddings-only inputs
    if req_state.prompt_token_ids is not None:
        input_tokens = req_state.prompt_token_ids
    elif req_state.prompt_embeds is not None:
        # For text-only embeddings, dummy token IDs are safe since
        # get_mrope_input_positions only uses len(input_tokens) when mm_features is empty
        seq_len = req_state.prompt_embeds.shape[0]
        input_tokens = list(range(seq_len))
        # Verify no mm_features remain (should be true after prompt_embeds filter)
        assert len(mrope_features) == 0, (
            "M-RoPE with prompt_embeds-only input should have no multimodal features"
        )
    else:
        raise ValueError(
            "M-RoPE requires either prompt_token_ids or prompt_embeds."
        )

    req_state.mrope_positions, req_state.mrope_position_delta = (
        mrope_model.get_mrope_input_positions(
            input_tokens,
            mrope_features,
        )
    )

Technically, in isolation, this method still crashes in the case where req_state.prompt_token_ids is None and req_state.mm_features, so the solution above still leaves that potential vector open. As far as can be determined, however, such a req_state is impossible in the first place in online mode, because it would require a /v1/completions request with prompt_embeds AND multimodal features, but the /v1/completions request schema does not expose multimodal inputs in any discernible way. Today, those are the only two endpoints with prompt_embeds support.

When in offline mode, it is technically possible to directly create an EngineCoreRequest that has prompt_embeds and not prompt_token_ids and mm_features, and pass that to LLM.generate. That would trigger this same assertion, and no validation would prevent that combination. It is strongly suspected, though, that this combination would be undefined in any model that support M-RoPE, because it would not be possible to determine which token positions correspond to mm_features. The proposed solution above would end up not setting req_state.mrope_positions and req_state.mrope_position_delta in this scenario, which could result in undefined behavior.

prompt_embeds is far more familiar here than M-RoPE, and it is understood that each model that supports it is responsible for defining its own get_mrope_input_positions which have varying implementations. There is insufficient knowledge to be prescriptive in how the two features should interact in the offline case, other than possibly raising a validation error earlier on preventing that combination (which would emulate the current assertion behavior). Regardless, in offline mode, the chances of a remote user being able to exploit this are slim-to-nil compared to the online case which is incredibly straightforward.

Impact

What kind of vulnerability is it? Who is impacted?

  • Denial of Service caused by an incorrect assertion inside of the GPUModelRunner which causes a fatal EngineCore exception
  • Any configuration with --enable-prompt-embeds and M-RoPE-supported model is vulnerable
  • The attack is extremely easy from the remote attacker's perspective (copying the official prompt_embeds online mode docs examples almost-verbatim, accounting for model-name and connection details, of course, will induce a guaranteed shutdown)

References

@jperezdealgaba jperezdealgaba published to vllm-project/vllm Jul 2, 2026
Published by the National Vulnerability Database Jul 6, 2026
Published to the GitHub Advisory Database Jul 20, 2026
Reviewed Jul 20, 2026
Last updated Jul 20, 2026

Severity

High

CVSS overall score

This score calculates overall vulnerability severity from 0 to 10 and is based on the Common Vulnerability Scoring System (CVSS).
/ 10

CVSS v4 base metrics

Exploitability Metrics
Attack Vector Network
Attack Complexity Low
Attack Requirements None
Privileges Required Low
User interaction None
Vulnerable System Impact Metrics
Confidentiality None
Integrity None
Availability High
Subsequent System Impact Metrics
Confidentiality None
Integrity None
Availability None

CVSS v4 base metrics

Exploitability Metrics
Attack Vector: This metric reflects the context by which vulnerability exploitation is possible. This metric value (and consequently the resulting severity) will be larger the more remote (logically, and physically) an attacker can be in order to exploit the vulnerable system. The assumption is that the number of potential attackers for a vulnerability that could be exploited from across a network is larger than the number of potential attackers that could exploit a vulnerability requiring physical access to a device, and therefore warrants a greater severity.
Attack Complexity: This metric captures measurable actions that must be taken by the attacker to actively evade or circumvent existing built-in security-enhancing conditions in order to obtain a working exploit. These are conditions whose primary purpose is to increase security and/or increase exploit engineering complexity. A vulnerability exploitable without a target-specific variable has a lower complexity than a vulnerability that would require non-trivial customization. This metric is meant to capture security mechanisms utilized by the vulnerable system.
Attack Requirements: This metric captures the prerequisite deployment and execution conditions or variables of the vulnerable system that enable the attack. These differ from security-enhancing techniques/technologies (ref Attack Complexity) as the primary purpose of these conditions is not to explicitly mitigate attacks, but rather, emerge naturally as a consequence of the deployment and execution of the vulnerable system.
Privileges Required: This metric describes the level of privileges an attacker must possess prior to successfully exploiting the vulnerability. The method by which the attacker obtains privileged credentials prior to the attack (e.g., free trial accounts), is outside the scope of this metric. Generally, self-service provisioned accounts do not constitute a privilege requirement if the attacker can grant themselves privileges as part of the attack.
User interaction: This metric captures the requirement for a human user, other than the attacker, to participate in the successful compromise of the vulnerable system. This metric determines whether the vulnerability can be exploited solely at the will of the attacker, or whether a separate user (or user-initiated process) must participate in some manner.
Vulnerable System Impact Metrics
Confidentiality: This metric measures the impact to the confidentiality of the information managed by the VULNERABLE SYSTEM due to a successfully exploited vulnerability. Confidentiality refers to limiting information access and disclosure to only authorized users, as well as preventing access by, or disclosure to, unauthorized ones.
Integrity: This metric measures the impact to integrity of a successfully exploited vulnerability. Integrity refers to the trustworthiness and veracity of information. Integrity of the VULNERABLE SYSTEM is impacted when an attacker makes unauthorized modification of system data. Integrity is also impacted when a system user can repudiate critical actions taken in the context of the system (e.g. due to insufficient logging).
Availability: This metric measures the impact to the availability of the VULNERABLE SYSTEM resulting from a successfully exploited vulnerability. While the Confidentiality and Integrity impact metrics apply to the loss of confidentiality or integrity of data (e.g., information, files) used by the system, this metric refers to the loss of availability of the impacted system itself, such as a networked service (e.g., web, database, email). Since availability refers to the accessibility of information resources, attacks that consume network bandwidth, processor cycles, or disk space all impact the availability of a system.
Subsequent System Impact Metrics
Confidentiality: This metric measures the impact to the confidentiality of the information managed by the SUBSEQUENT SYSTEM due to a successfully exploited vulnerability. Confidentiality refers to limiting information access and disclosure to only authorized users, as well as preventing access by, or disclosure to, unauthorized ones.
Integrity: This metric measures the impact to integrity of a successfully exploited vulnerability. Integrity refers to the trustworthiness and veracity of information. Integrity of the SUBSEQUENT SYSTEM is impacted when an attacker makes unauthorized modification of system data. Integrity is also impacted when a system user can repudiate critical actions taken in the context of the system (e.g. due to insufficient logging).
Availability: This metric measures the impact to the availability of the SUBSEQUENT SYSTEM resulting from a successfully exploited vulnerability. While the Confidentiality and Integrity impact metrics apply to the loss of confidentiality or integrity of data (e.g., information, files) used by the system, this metric refers to the loss of availability of the impacted system itself, such as a networked service (e.g., web, database, email). Since availability refers to the accessibility of information resources, attacks that consume network bandwidth, processor cycles, or disk space all impact the availability of a system.
CVSS:4.0/AV:N/AC:L/AT:N/PR:L/UI:N/VC:N/VI:N/VA:H/SC:N/SI:N/SA:N

EPSS score

Exploit Prediction Scoring System (EPSS)

This score estimates the probability of this vulnerability being exploited within the next 30 days. Data provided by FIRST.
(30th percentile)

Weaknesses

Reachable Assertion

The product contains an assert() or similar statement that can be triggered by an attacker, which leads to an application exit or other behavior that is more severe than necessary. Learn more on MITRE.

CVE ID

CVE-2026-55514

GHSA ID

GHSA-33cg-gxv8-3p8g

Source code

Credits

Loading Checking history
See something to contribute? Suggest improvements for this vulnerability.