vLLM: Incomplete CVE-2025-62164 remediation can be bypassed by concurrent prompt parts

Executive SummaryThe follow-up protection for CVE-2025-62164 is incomplete at vLLM revision 26587f9519e22a5c4549ead7595ad9ca3229c4fd. It wraps serialized prompt-embedding reconstruction and dense conversion in torch.sparse.checksparsetensorinvariants, but PyTorch 2.11.0 implements that context wi...

PYSEC-2026-3938 vLLM: Incomplete CVE-2025-62164 remediation can be bypassed by concurrent prompt parts

Executive Summary The follow-up protection for CVE-2025-62164 is incomplete at vLLM revision 26587f9519e22a5c4549ead7595ad9ca3229c4fd. It wraps serialized prompt-embedding reconstruction and dense conversion in torch.sparse.checksparsetensorinvariants, but PyTorch 2.11.0 implements that context...

vLLM: Incomplete CVE-2025-62164 remediation can be bypassed by concurrent prompt parts

Executive Summary The follow-up protection for CVE-2025-62164 is incomplete at vLLM revision 26587f9519e22a5c4549ead7595ad9ca3229c4fd. It wraps serialized prompt-embedding reconstruction and dense conversion in torch.sparse.checksparsetensorinvariants, but PyTorch 2.11.0 implements that context...

CVE-2026-73557 vLLM: Incomplete CVE-2025-62164 remediation can be bypassed by concurrent prompt parts

vLLM is an inference and serving engine for large language models. From 0.20.2rc0 until 0.26.0, safeloadpromptembeds in vllm/renderers/embedutils.py uses torch.sparse.checksparsetensorinvariants, whose process-global save, enable, and restore state can be raced by concurrent promptembeds parts...

CVE-2026-73557

CVE-2026-73557 affects vLLM between 0.20.2rc0 and 0.26.0. The flaw arises in safe_load_prompt_embeds in vllm/renderers/embed_utils.py, where using torch.sparse.check_sparse_tensor_invariants with a process-global save/enable/restore state can be raced via concurrent prompt_embeds (POST /v1/chat/c...

Denial Of Service (DoS)

vLLM is vulnerable to Denial of Service DoS. The vulnerability is due to improper handling of prompt embedding payloads for models using M-RoPE, which allows an authenticated attacker to trigger an assertion failure and crash the entire server by sending a crafted /v1/completions request...

The Trojan Example: Jailbreaking LLMs through Template Filling and Unsafety Reasoning

Large Language Models LLMs have advanced rapidly and now encode extensive world knowledge. Despite safety fine-tuning, however, they remain susceptible to adversarial prompts that elicit harmful content. Existing jailbreak techniques fall into two categories: white-box methods e.g., gradient-base...