vLLM before 0.33.0: one request can crash EngineCore or exhaust the frontend, and a malicious model config can swap in another local model (CVE-2026-130684/130685/130686)
Three vLLM CVEs fixed in 0.33.0 (released 2 November 2026) were published on 18 November. With --skip-tokenizer-init, one unauthenticated request carrying an out-of-range prompt token id crashes the shared EngineCore for every tenant. Any multimodal deployment can be hit by a single request listing many remote image URLs, all downloaded and decoded before length checks. And a served model whose config.json has speculators_config.verifier.name_or_path can redirect vLLM to another local checkpoint and leak its tokenizer via /detokenize.
What it is
CVE-2026-130685 / GHSA-q9v9-qpjp-fc24 (vLLM 0.28.0 to before 0.33.0): the vocab-range check added for GHSA-25q3-v2hm-8vpf only runs when a tokenizer exists, so with --skip-tokenizer-init a large token id in /v1/completions (and /v1/embeddings, /pooling) reaches the embedding as an index and EngineCore shuts down. The OpenAI-compatible server has no auth by default.
CVE-2026-130686 / GHSA-5r3c-5x4v-8gjh (0.6.0 to before 0.33.0): there is no request-wide byte or memory budget for media, and the unspecified per-modality count defaults to 999. One compact request to /v1/messages or non-Harmony /v1/responses can tie up outbound connections, decoder CPU and API-frontend memory.
CVE-2026-130684 / GHSA-qrj2-j723-m6w8 (before 0.33.0): maybe_override_with_speculators runs for every non-cloud-storage model and takes speculators_config verifier.name_or_path from the artifact's config.json as both model and tokenizer with no containment check, even with trust_remote_code off. A malicious model publisher can point it at another local checkpoint directory.
How to check
- Run pip show vllm (or check your image tag) and compare with 0.33.0.
- Extra exposure: --skip-tokenizer-init deployments (token-in/token-out, RL serving), multimodal models that accept remote image URLs, and servers that load models from third-party repositories.
How to fix
- Upgrade vLLM to 0.33.0 or later.
- Put vLLM behind an authenticating gateway with request-size and rate limits; set --limit-mm-per-prompt to the image count you actually need.
- Only serve model repositories you trust, and check config.json for a speculators_config key before serving a third-party model.
pip show vllm | grep Version
# does a downloaded model try to redirect the verifier?
grep -n speculators_config /path/to/model/config.json Sources
- GitHub advisory GHSA-q9v9-qpjp-fc24: unauthenticated EngineCore crash via out-of-range prompt token id ↗
- GitHub advisory GHSA-5r3c-5x4v-8gjh: missing request-wide media budget ↗
- GitHub advisory GHSA-qrj2-j723-m6w8: speculators config verifier.name_or_path directory traversal ↗
- NVD: CVE-2026-130685 ↗
- NVD: CVE-2026-130686 ↗
- NVD: CVE-2026-130684 ↗
- highCVE-2026-115912: vLLM 0.32.0 fixes a forged shared-memory handle that reaches pickle.loads, EPS images that can reach Ghostscript, and five ways to stall or crash the server
- mediumvLLM: five more request-driven DoS and info-leak CVEs (CVE-2026-126394, 126395, 126396, 126398, 126399), fixed in 0.30.0 / 0.31.0
- criticalCVE-2026-61732: web pages can forge operator turns in the Decepticon hacking agent with ChatML tokens and run commands, upgrade to 1.1.17
- criticalCVE-2026-22778: vLLM has RCE In Video Processing
- criticalCVE-2025-47277: vLLM Allows Remote Code Execution via PyNcclPipe Communication Service
- criticalCVE-2025-32444: vLLM Vulnerable to Remote Code Execution via Mooncake Integration