CVE-2026-93989
vLLM through 0.29.0 Cross-Request Logits Corruption via bad_words
No known exploitation. EPSS puts it in the 22nd percentile. No fix published yet.
What to do
No fixed build or workaround is published yet. Limit exposure and watch for a patch.
What it is
From the CVE record
vLLM through 0.29.0 fails to properly validate bad_words token indices against the model's generation output width in SamplingParams.update_from_tokenizer(). Attackers can supply out-of-bounds token indices that corrupt logits memory of concurrent requests, causing different in-flight HTTP requests to return incorrect tokens.
In plain language
No plain-language summary for this CVE yet.
Exploitation
Where each signal puts this CVE on the scale from published to confirmed exploited.
- CISA KEV
Not in the catalog. CISA has not confirmed exploitation.
- Public exploits
No public exploit or proof of concept found in the sources we track.
- EPSS
0.3% chance of exploitation activity in the next 30 days, which ranks it in the 22nd percentile of scored CVEs.
Exploit Prediction Scoring System, FIRST.org. A probability, not a confirmation.
Lifecycle
5 events over 3 days, from the signal feeds we watch.
- Record updatedrecord updated
- Publishedweakness classified, record updated
Affected products
Technical detail
CVSS 3.1 vector
Open in the CVSS calculatorCVSS:3.1/AV:N/AC:H/PR:L/UI:N/S:U/C:N/I:L/A:N
Scored 3.1 by NVD.
How it is reached
- Attack Vector NetworkExploitable remotely over the network without any special conditions
- Attack Complexity HighRequires specific conditions like a race condition or non-default configuration
- Privileges Required LowRequires basic user-level privileges
- User Interaction NoneNo user interaction needed — fully automated exploitation
Scope
- Scope UnchangedImpact is limited to the vulnerable component itself
Impact if exploited
- Confidentiality NoneNo confidentiality impact
- Integrity LowData modification is possible but limited in scope or consequence
- Availability NoneNo availability impact
Weaknesses
Sources
References in the record
- github.com/vllm-project/vllm/pull/48824
- github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/vllm/v1/worker/gpu/sample/bad_words.py
- github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/vllm/sampling_params.py
And 2 more references. See all after sign-in
Watch the software you run.
My Stack ranks new CVEs for your products by real-world exploitation, so the next serious one reaches you without reading every advisory.
We'll flag the next CVE, public exploit or patch for Vllm, not every advisory.
A free account adds
- The full version matrix and every affected product
- Exploit links, proofs of concept and Metasploit modules
- Email alerts for the products you watch
- The same data over REST API, MCP and CLI