Database/AI/ML frameworks & serving
vLLM: structured-output request failures escape request scope and terminate the shared engine
Impact
Failures in constrained-generation handling reach EngineCore's fatal-error path instead of failing just the request. Three routes are named: a per-request backend mismatch re-raises a grammar compilation exception; padding produced by the ngram_gpu speculative-decoding mode passes a negative token into guidance validation; and the Rust frontend admits empty structured-output values that the Python frontend rejects. Any of these lets an ordinary-looking JSON-schema or grammar request kill the serving process, dropping all concurrent tenants on that replica and forcing a weight reload. Structured output is heavily used by agent and tool-calling workloads, so this is reachable in normal traffic, not only by a deliberate attacker.
Who can reach it
Any authenticated client that can send a structured-output (guided decoding / JSON schema / grammar) request over the network, CVSS AV:N/PR:L. No special configuration is needed for the frontend-validation route; the speculative-decoding route requires ngram_gpu to be enabled.
What to do
Upgrade vLLM to 0.30.0 and restart the serving replicas, rolling them to keep the endpoint available. As an interim measure, disable ngram_gpu speculative decoding and validate structured-output fields at the gateway; neither closes the grammar-backend route, so the upgrade is the real fix.
References
Related entries
- MLflow: unauthenticated scorer/invoke deserialization lets a caller drive outbound requests to internal hostsCVE-2026-115886 · MLflow scorer endpoints (Scorer.model_validate_json deserialization)Medium
- vLLM: unbounded chat-template rendering lets one request wedge or exhaust the serving processCVE-2026-115911 · vLLM OpenAI-compatible server (request-supplied Jinja chat template rendering)Medium
- vLLM: Phi-4 audio encoder materializes quadratic attention matrices, letting one long clip OOM the engineCVE-2026-115913 · vLLM Phi-4 multimodal audio encoder (phi4mm_audio.py eager attention)Medium
- vLLM: an all-zero image on the DeepSeek-OCR path trips a fatal assertion and shuts the engine downCVE-2026-115914 · vLLM DeepSeek-OCR image path (_parse_and_validate_image_input / sanity_check_mm_encoder_outputs)Medium
- vLLM: off-by-one prompt-token-ID bound lets a request crash the inference engineCVE-2026-115916 · vLLM tokenizer input validation (vllm/tokenizers/hf.py max_token_id)Medium
- vLLM: request-supplied chat template bypasses trust_request_chat_template and exhausts the serverCVE-2026-115918 · vLLM /v1/responses endpoint (chat_template_kwargs trust policy)Medium
This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.