What happened
vLLM, a widely-deployed LLM inference and serving engine, disclosed five CVEs on 2026-08-13 affecting versions from 0.19.0–0.27.0: CVE-2026-73559 (CVSS 6.5) unbounded prompt list in CompletionRequest.prompt causing resource exhaustion; CVE-2026-73557 (CVSS 6.3) a race condition in safe_load_prompt_embeds via process-global torch.sparse state; CVE-2026-73556 (CVSS 5.3) unvalidated structured_outputs.regex passed to RegexParser without timeout, enabling ReDoS; CVE-2026-73555 (CVSS 5.3) unsanitized exception messages leaking traceback information; CVE-2026-73558 (CVSS 5.3) an integer overflow in a CUDA activation kernel that can leak another batched user's input data.
Why it matters
vLLM is one of the most widely deployed open-source inference engines for production LLM serving; the batch-processing information leak (CVE-2026-73558) is particularly notable because it can cause one tenant's request to receive fragments of another user's input in a shared inference batch — a cross-tenant data leakage risk in multi-tenant GPU inference hosting, alongside multiple unauthenticated DoS vectors against the OpenAI-compatible API.
Attack vector
Unauthenticated requests to the OpenAI-compatible /v1/completions endpoint with oversized prompt lists, crafted regex in structured-outputs requests, or race conditions during concurrent prompt-embeds processing; batch integer overflow triggered by request timing/sizing in the same inference batch as a victim
Affected systems
vLLM 0.19.0–0.27.0 (varies per CVE)
Mitigation
Upgrade to vLLM 0.26.0 or 0.27.0 depending on the specific CVE per vendor commits