
Cloud Vulnerability DB
A community-led vulnerabilities database
CVE-2026-112270 is an uncontrolled memory allocation vulnerability in SGLang's Mooncake disaggregation backend that allows unauthenticated network-adjacent attackers to exhaust Prefill scheduler memory, causing an OOM-kill. It affects SGLang versions through 0.5.22 and was disclosed on October 19, 2026. The vulnerability resides in the handle_staging_rsp() function in staging_handler.py. It carries a CVSS v3.1 base score of 6.5 (Medium) and a CVSS v4.0 base score of 7.1 (High) (Github Advisory, Feedly).
The root cause is CWE-770 (Allocation of Resources Without Limits or Throttling) in StagingTransferInfo.set_chunk(), which grows three Python lists (offsets, rounds, ends) to chunk_idx + 1 entries using an unbounded while loop with no maximum index check, per-peer quota, or global quota (Github Issue). The Prefill receive thread in MooncakeKVManager.start_prefill_thread() (in conn.py) dispatches any ZeroMQ message whose first frame is STAGING_RSP directly to handle_staging_rsp() without authenticating the sender, verifying room ownership, or validating the session binding. An attacker first sends a crafted eight-frame transfer message to create a fake room and session in self.transfer_infos, then sends a STAGING_RSP message with an arbitrarily large chunk_idx value; the dispatch path is active even when SGLANG_DISAGG_STAGING_BUFFER=0 (staging disabled). The Prefill bootstrap /route endpoint exposes the rank IP and port without authentication, enabling any network-adjacent peer to obtain the ZeroMQ socket address needed to initiate the attack (Github Source, Github Source).
Successful exploitation causes the SGLang Prefill scheduler process to be terminated by the Linux OOM killer, rendering the LLM inference service completely unavailable. In a validated two-host test on SGLang 0.5.15.post1, 3,160 unique sessions with chunk_idx=200000 increased Prefill cgroup memory by 12.23 GiB, slowed concurrent inference requests by 46.9× on average, and caused the scheduler to be OOM-killed; a single 85-byte message pair with chunk_idx=500000 grew three internal lists from 1 to 500,001 entries in 0.08 seconds (Github Issue). There is no confidentiality or integrity impact — the vulnerability is a pure availability (denial-of-service) issue with no lateral movement potential beyond disrupting the targeted Prefill node (Github Advisory).
A functional proof-of-concept Python script (poc.py) was published as part of the bug report, requiring only pyzmq and network adjacency to the Prefill bootstrap port; no authentication or prior application state is needed (Github Issue). As of the disclosure date, there is no evidence of in-the-wild exploitation or threat actor attribution, and the NVD SSVC assessment lists exploitation as "none" (Feedly). The EPSS score is approximately 0.249% (16th percentile), indicating a low but non-negligible probability of exploitation in the near term (Github Advisory). The vulnerability is not listed in the CISA KEV catalog.
/route HTTP endpoint (e.g., http://<PREFILL_IP>:<BOOTSTRAP_PORT>/route?prefill_dp_rank=0&prefill_cp_rank=0&target_tp_rank=0&target_pp_rank=0) to obtain the rank_ip and rank_port of the Prefill ZeroMQ PUSH socket.pyzmq, create a zmq.PUSH socket and connect it to tcp://<rank_ip>:<rank_port> obtained from the /route response.transfer_infos dictionary — no real Decode worker or prior inference request is required.STAGING_RSP multipart message with the same room and session IDs but with a very large chunk_idx value (e.g., 200,000 or 500,000). The Prefill receive thread dispatches this to handle_staging_rsp(), which calls StagingTransferInfo.set_chunk(), growing three Python lists to chunk_idx + 1 entries.STAGING_RSP injection across thousands of unique room IDs to multiply memory consumption. Each iteration allocates a new StagingTransferInfo object with lists of the target size.chunk_idx=200,000), the Prefill scheduler process exhausts available memory and is terminated by the Linux OOM killer, taking the inference service offline (Github Issue)./route endpoint from hosts that are not registered Decode workers; high-frequency ZeroMQ PUSH connections to the Prefill rank port from unknown source IPs.memory.current without a corresponding increase in active inference requests; memory.current approaching the configured MemoryMax cgroup limit.A process of this unit has been killed by the OOM killer for the Prefill scope; SGLang log warnings like STAGING_RSP RECV but tinfo=None room=<N> chunk=<large_N> session=<UUID> for rooms that were not legitimately created.oom_kill_process) referencing the SGLang scheduler process (Github Issue).Upgrade SGLang to a version newer than 0.5.22, which includes the fix (commit 2b3cb36) that drops STAGING_RSP messages for chunks the engine never requested (Github Issue). As an immediate workaround, restrict network access to the Prefill bootstrap HTTP port and ZeroMQ rank port to trusted network segments only (e.g., via firewall rules or network policy), preventing unauthenticated peers from reaching these endpoints. Additionally, implement input validation on chunk_idx values in STAGING_RSP messages and configure memory alerts on Prefill cgroup memory.current to detect abnormal allocation patterns early (Feedly).
The vulnerability was responsibly reported by researcher yyymk via the SGLang GitHub issue tracker on October 18, 2026, with a detailed write-up including reproduction steps, measured memory impact, and a working PoC script (Github Issue). SGLang maintainer he-yufeng acknowledged the issue and committed a fix (commit 2b3cb36) on October 22, 2026, with the fix targeting the main branch. The reporter noted this is the "opposite direction" of the previously disclosed CVE-2026-93838, which covers the Decode-side STAGING_REQ path, emphasizing that fixing one path does not remediate the other.
Source: This report was generated using AI
Free Vulnerability Assessment
Evaluate your cloud security practices across 9 security domains to benchmark your risk level and identify gaps in your defenses.
Get a personalized demo
"Best User Experience I have ever seen, provides full visibility to cloud workloads."
"Wiz provides a single pane of glass to see what is going on in our cloud environments."
"We know that if Wiz identifies something as critical, it actually is."