CVE-2026-115916VulnerabilityPublished 28 Oct 2026

vLLM: Out-of-vocabulary prompt token id passes input validation and triggers a CUDA device-side assert (engine crash /

Affects vllmSource GCVE Database

Authoritative record:CVE-2026-115916· Medium 6.5

Not known to be exploited · CVSS 6.5 Medium : upgrade to 0.32.0 when convenient.

Data as of 28 Oct 2026, 22:12 UTC

CVSS 3.1 base score

6.5Medium

How severe the flaw is on the 0–10 CVSS scale.

Patch: Available – fixed in 0.32.0

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

Bold values are the worst case: easiest to exploit or largest impact.

Attack Vector: Network: The vulnerable component is bound to the network stack and the set of possible attackers extends beyond the other options listed, up to and including the entire Internet. Such a vulnerability is often termed 'remotely exploitable' and can be thought of as an attack being exploitable at the protocol level one or more network hops away (e.g., across one or more routers).
Attack Complexity: Low: Specialized access conditions or extenuating circumstances do not exist. An attacker can expect repeatable success against the vulnerable component.
Privileges Required: Low: The attacker is authorized with (i.e., requires) privileges that provide basic user capabilities that could normally affect only settings and files owned by a user. Alternatively, an attacker with Low privileges may have the ability to cause an impact only to non-sensitive resources.
User Interaction: None: The vulnerable system can be exploited without any interaction from any user.
Scope: Unchanged: An exploited vulnerability can only affect resources managed by the same security authority. In this case, the vulnerable component and the impacted component are either the same, or both are managed by the same security authority.
Confidentiality: None: There is no loss of confidentiality within the impacted component.
Integrity: None: There is no loss of integrity within the impacted component.
Availability: High: There is total loss of availability, resulting in the attacker being able to fully deny access to resources in the impacted component; this loss is either sustained (while the attacker continues to deliver the attack) or persistent (the condition persists even after the attack has completed). Alternatively, the attacker has the ability to deny some availability, but the loss of availability presents a direct, serious consequence to the impacted component (e.g., the attacker cannot disrupt existing connections, but can prevent new connections; the attacker can repeatedly exploit a vulnerability that, in each instance of a successful attack, leaks a only small amount of memory, but after repeated exploitation causes a service to become completely unavailable).

Definitions: FIRST CVSS v3.1 specification.

Exploitation evidence

Disclosed: evidence found.
CVE published 26 Oct 2026
Detectable: no evidence on record.
No public nuclei template indexed
Weaponised: no evidence on record.
No public exploit code indexed in Metasploit or Exploit-DB
Exploited in the wildExploited: no evidence on record.
Not in CISA KEV and no source reports exploitation
Observed by OffSeq MirageSeen by Mirage: no evidence on record.
No attack traffic recorded by OffSeq Mirage sensors

A missing step is not evidence of safety. Sources: cve.org (opens in a new tab), the Metasploit, Exploit-DB and nuclei-templates index as of 26 Oct 2026.

Disclosed: The CVE record is public on cve.org. Its publication is t = 0.
Detectable: A public nuclei template can find exposed instances, for defenders and attackers alike.
Weaponised: Public exploit code exists in Metasploit or Exploit-DB, or CISA SSVC or the CVSS exploit maturity reports a public proof of concept.
Exploited in the wild: CISA KEV lists it, CISA SSVC rates exploitation active, or a source reports attacks.
Observed by OffSeq Mirage: OffSeq Mirage honeypot sensors recorded attack traffic for this CVE, or found exposed hosts without attack traffic.

Moments are spaced evenly, not to scale; a dashed line marks a gap of six months or more. Deltas count from t = 0. Sources that give only a date (CISA KEV, Exploit-DB, or a time of exactly 00:00Z) read “same day” rather than hours. Metasploit modules carry the flaw’s disclosure date, not their release date, and nuclei templates carry no date, so both appear above but not on the timeline.

What to do

Upgrade vllm to 0.32.0.

Affected software

Product	Affected	Fixed in
PyPIvllm	Affected <0.32.0	Fixed in 0.32.0

Check your own hosts for these packages with threat-finder (opens in a new tab), our free open-source scanner.

Description

For a model whose vocabulary size is an exact multiple of the embedding padding size (64) — e.g. any Llama-2-family model, vocab 32000 — a single request carrying a pre-tokenized prompt with token id equal to the vocab size (32000) passes vLLM's input-id validation but is one past the last embedding row. At tensor-parallel size 1 the embedding lookup does not mask indices, so the id reaches `F.embedding` out of range and raises a CUDA device-side assert, which poisons the engine's CUDA context and takes the engine down. Any client that can send pre-tokenized input (`/v1/completions` with `prompt` as token ids, `/tokenize`-fed flows, or the `kv_transfer_params["prompt_token_ids"]` reuse path) can trigger it. ## Root cause `vllm/v1/engine/input_processor.py` accepts a token id when `id <= max(tokenizer.max_token_id, model_config.get_vocab_size() - 1)`. `vllm/tokenizers/hf.py` computes `max_token_id = max(max(vocab.values()), tokenizer.vocab_size)`. For Llama-2 tokenizers `max(vocab.values()) = 31999` but `tokenizer.vocab_size = 32000`, so `max_token_id = 32000` — one greater than the largest valid id. The engine bound therefore becomes `max(32000, 31999) = 32000`, and id 32000 is accepted. The model's embedding has `pad_vocab_size(32000, 64) = 32000` rows (no padding), valid indices 0..31999. At `VocabParallelEmbedding.forward` with `tp_size == 1` there is no masking, so index 32000 hits `F.embedding` out of range. ## Affected - tp=1 (single GPU per model — common per-copy layout). tp>1 masks via `get_masked_input_and_mask`, so the id is clamped, not an OOB — likely wrong output rather than a crash. - Models whose vocab is a multiple of 64 with `tokenizer.vocab_size` one past the max real id (Llama-2 family, vocab 32000, is the clear case). Qwen2.5 is NOT affected (padded headroom). ## Suggested fix max_token_id in vllm/tokenizers/hf.py now uses tokenizer.vocab_size - 1, so an id equal to the vocabulary size fails input validation instead of reaching the embedding. Fixed in [MR59491](https://github.com/vllm-project/vllm/pull/59491), released in v0.32.0. ## Reporter Shijie Lyu (GitHub shijie-lyu lshjhf@gmail.com). Reported privately per SECURITY. This advisory was fixed in [MR59491](https://github.com/vllm-project/vllm/pull/59491).

CWE-129Improper Validation of Array Index (opens in a new tab)
Variant weaknessLikelihood of exploit: HighOWASP Top 10 A05:2025: Injection (opens in a new tab)
The product uses untrusted input when calculating or using an array index, but the product does not validate or incorrectly validates the index to ensure the index references a valid position within the array.
Architecture and Design · Input Validation
Use an input validation framework such as Struts or the OWASP ESAPI Validation API. Note that using a framework does not automatically address all input validation problems; be mindful of weaknesses that could arise from misusing the framework itself (CWE-1173).
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…
Requirements · Language Selection
Use a language that does not allow this weakness to occur or provides constructs that make this weakness easier to avoid.
OWASP Cheat Sheets
XML Security (opens in a new tab)
Mass Assignment (opens in a new tab)
Injection Prevention (opens in a new tab)
SQL Injection Prevention (opens in a new tab)
Injection Prevention in Java (opens in a new tab)
OWASP ASVS 4.0.3
V13.3.1 (opens in a new tab)
V5.1.4 (opens in a new tab)
V5.1.3 (opens in a new tab)
Attack patterns
Overflow Buffers (opens in a new tab)
From MITRE CWE, MITRE CAPEC and OWASP (Top 10, OpenCRE).
CWE-193Off-by-one Error (opens in a new tab)
Base weakness
A product calculates or uses an incorrect maximum or minimum value that is 1 more, or 1 less, than the correct value.
Implementation
When copying character arrays or using character manipulation methods, the correct size parameter must be used to account for the null terminator that needs to be added at the end of the array.
From MITRE CWE, MITRE CAPEC and OWASP (Top 10, OpenCRE).

References

Community reviews

Loading reviews…

Exploitation evidence

What to do

Affected software

Description

WeaknessesCWE-129, CWE-193

Technical details

References

Community reviews

Related threats