CVEs
Vulnerabilities discovered through Faultline's analysis of LLM inference systems. Each entry links to the public advisory and any accompanying writeup.
CVE-2026-10300low · 3.4DisclosedSGLang LoRA manager reachable assertion via lora_path triggers remote DoS
Affects SGLang 0.5.10.post1 · Disclosed 2026-06-01
A reachable assertion in SGLang's LoRA manager (python/sglang/srt/lora/lora_manager.py) within the inference HTTP endpoint. Manipulating the lora_path argument triggers an assertion failure, producing a remote denial-of-service condition. Remotely exploitable; a public exploit exists.
Credit: Yunze Zhao
sglanglorareachable-assertiondenial-of-servicecwe-617CVE-2026-9540medium · 4.8DisclosedvLLM n_completions + logprobs decode amplification causes TTFT DoS across co-scheduled requests
Affects vLLM 0.19.0 · Disclosed 2026-05-26
A denial-of-service flaw in vLLM's OpenAI-compatible serving path: a single request combining n_completions=N with logprobs=K forces every co-scheduled request on the same batch to pay an N x full-vocabulary softmax/sort cost per decode step. Victim requests experienced 76x-423x time-to-first-token regressions. Remotely triggerable, no auth or user interaction.
Credit: Yunze Zhao
vllmopenai-apidecode-amplificationresource-exhaustiondenial-of-servicecwe-404CVE-2026-7141medium · 5.1DisclosedvLLM KV-cache block recycle leaves stale K/V, causing non-deterministic output and NaN propagation
Affects vLLM ≤ 0.19.0 · Disclosed 2026-04-27
In vLLM's V1 engine, the KV-block zeroing pipeline (introduced in #35219) is gated by has_mamba_layers in vllm/v1/kv_cache_interface.py and therefore only runs for Mamba models. For FullAttention models, KV blocks returned to the free pool on request completion or cancellation are reused without zeroing, so a subsequent request can decode from stale K/V belonging to a previous request. This produces non-deterministic output at temperature=0 (no prefix caching required) and can propagate NaNs through masked softmax in partial-block tail slots.
Credit: Yunze Zhao
vllmkv-cacheuninitialized-memorynon-determinismcwe-908