CacheTrap

CacheTrap Repository for "CacheTrap: Unveiling a Stealthier Gray-Box Trojan against LLMs" , IEEE/ACM International Conference on Computer-Aided Design ICCAD, 2026. arXiv CacheTrap searches for a single bit flip in the KV cache of a fine-tuned LLM classifier and measures the resulting attack succe...

Server-side Request Forgery (SSRF)

Overview vllm is an A high-throughput and memory-efficient inference and serving engine for LLMs Affected versions of this package are vulnerable to Server-side Request Forgery SSRF via the P2pNcclConnector KV cache transfer connector, which exposes KV cache contents to network-accessible peers...

CVE-2026-116176 vLLM 0.9.2 before 0.24.0 SSRF via X-Request-Id in P2pNcclConnector

vLLM from 0.9.2 before 0.24.0 contains a server-side request forgery vulnerability that allows API callers to force outbound connections by embedding a decode address in the X-Request-Id header. Attackers can make a P2pNcclConnector prefill worker ZMQ-connect to an attacker host, leaking the NCCL...

CVE-2026-116176 vLLM 0.9.2 before 0.24.0 SSRF via X-Request-Id in P2pNcclConnector

vLLM from 0.9.2 before 0.24.0 contains a server-side request forgery vulnerability that allows API callers to force outbound connections by embedding a decode address in the X-Request-Id header. Attackers can make a P2pNcclConnector prefill worker ZMQ-connect to an attacker host, leaking the NCCL...

CVE-2026-111392 SGLang through 0.5.21 KV Cache Poisoning via mm_hashes in /generate

SGLang through 0.5.21 trusts caller-supplied mmhashes values as the multimodal prefix-cache key, allowing remote attackers to poison cached KV entries across requests. Attackers can send a different image under a known or predicted hash via /generate to disclose another request's image content or...

CVE-2026-111392 SGLang through 0.5.21 KV Cache Poisoning via mm_hashes in /generate

SGLang through 0.5.21 trusts caller-supplied mmhashes values as the multimodal prefix-cache key, allowing remote attackers to poison cached KV entries across requests. Attackers can send a different image under a known or predicted hash via /generate to disclose another request's image content or...

CVE-2026-111392

SGLang versions through 0.5.21 are vulnerable to KV Cache Poisoning due to insufficient verification of mm_hashes values used as multimodal prefix-cache keys. A remote attacker can exploit this by sending a different image with a known or predicted hash via the /generate endpoint. This allows the...

CVE-2026-111392 SGLang through 0.5.21 KV Cache Poisoning via mm_hashes in /generate

SGLang through 0.5.21 trusts caller-supplied mmhashes values as the multimodal prefix-cache key, allowing remote attackers to poison cached KV entries across requests. Attackers can send a different image under a known or predicted hash via /generate to disclose another request's image content or...

Stealthy Poisoning Attacks Via Reused KV Cache of Reasoning

Modern large language model LLM systems reuse key-value KV caches to accelerate inference. Recent work has shown that reusing KV caches from reasoning traces can also improve task performance, similar to retrieval-augmented generation RAG. We demonstrate that this form of KV cache reuse introduce...

WS-2026-0166

vllm 0.9.2 - 0.23.0: P2pNcclConnector parses a host:port from the client-controlled X-Request-Id header and connects to it over ZMQ/NCCL, letting an API caller exfiltrate KV-cache tensors to an attacker host GHSA-vfp2-c8pq-v6h6...

Visual Memory Attacks Can Persist through the KV Cache

Modern language model systems operate autonomously over increasingly long contexts containing untrusted text and images. Can an adversarial input continue to steer a model even after that input is removed from its context? We show that attacks can be trained to persist through the key/value KV...

CVE-2026-103765: Missing Authentication for Critical Function

Mooncake through 0.3.13.post1 contains a missing authentication vulnerability in the HTTP metadata server /metadata handler that allows unauthenticated attackers to read, overwrite, and delete transfer engine metadata keys. Attackers can poison segment descriptors such as tcpdataport or re-create...

CVE-2026-103765

Mooncake through 0.3.13.post1 contains a missing authentication vulnerability in the HTTP metadata server /metadata handler that allows unauthenticated attackers to read, overwrite, and delete transfer engine metadata keys. Attackers can poison segment descriptors such as tcpdataport or re-create...

CVE-2026-103765 Mooncake through 0.3.13.post1 Missing Authentication in HTTP Metadata Server

Mooncake through 0.3.13.post1 contains a missing authentication vulnerability in the HTTP metadata server /metadata handler that allows unauthenticated attackers to read, overwrite, and delete transfer engine metadata keys. Attackers can poison segment descriptors such as tcpdataport or re-create...

CVE-2026-103764 Mooncake transfer engine before 0.3.13 Unauthenticated Arbitrary Memory Read/Write via TCP Transport

Mooncake transfer engine before 0.3.13 contains an untrusted pointer dereference in ServerSession::readHeader that allows unauthenticated attackers to read and write arbitrary process memory via the TCP transport data port. Attackers can send a crafted SessionHeader with arbitrary addr and size...

PT-2026-104057

Mooncake transfer engine before 0.3.13 contains an untrusted pointer dereference in ServerSession::readHeader that allows unauthenticated attackers to read and write arbitrary process memory via the TCP transport data port. Attackers can send a crafted SessionHeader with arbitrary addr and size...

SUSE CVE-2026-94627

vLLM Mooncake connector through 0.29.0 fails to properly manage GPU KV cache block ownership when concurrent child requests share a single transfer ID in prefill/decode disaggregated deployments. Attackers can trigger GPU memory exhaustion by submitting completion requests with multiple prompts,...

CVE-2026-94627

A flaw was found in vLLM. A remote attacker can exploit a vulnerability in the Mooncake connector's management of GPU Key-Value KV cache block ownership. By submitting completion requests with multiple prompts that share a single transfer ID, attackers can cause orphaned KV cache blocks to...

EUVD-2026-84221

vLLM Mooncake connector through 0.29.0 fails to properly manage GPU KV cache block ownership when concurrent child requests share a single transfer ID in prefill/decode disaggregated deployments. Attackers can trigger GPU memory exhaustion by submitting completion requests with multiple prompts,...

Improper Handling of Exceptional Conditions

Overview vllm is an A high-throughput and memory-efficient inference and serving engine for LLMs Affected versions of this package are vulnerable to Improper Handling of Exceptional Conditions in mooncakeconnector.py via the Mooncake KV transfer connector, when a remote KV cache load fails during...