Cyber Resilience

CVE-2026-3059

CriticalPublic PoCRCE

Published: 12 March 2026

Published
12 March 2026
Modified
07 April 2026
KEV Added
Patch
CVSS Score v3.1 9.8 CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H
EPSS Score 0.0153 71.6th percentile
Risk Priority 70 floored blend · peak EPSS

Summary

CVE-2026-3059 is a critical-severity Deserialization of Untrusted Data (CWE-502) vulnerability in Lmsys Sglang. Its CVSS base score is 9.8 (Critical).

Operationally, exploitation aligns with the MITRE ATT&CK technique Exploit Public-Facing Application (T1190); ranked in the top 28.4% of CVEs by exploit likelihood; it is not currently listed in the CISA KEV catalog; a public proof-of-concept is referenced.

The strongest mitigations our analysis identified are NIST 800-53 AC-3 (Access Enforcement) and SI-10 (Information Input Validation).

Deeper analysis

SGLang's multimodal generation module is vulnerable to unauthenticated remote code execution. The flaw exists in the ZMQ broker, which calls pickle.loads() on data received without authentication or validation, allowing deserialization of untrusted content under CWE-502.

An attacker with network access can supply a malicious serialized payload to the broker and obtain arbitrary code execution, resulting in complete compromise of confidentiality, integrity, and availability on the host.

The GitHub security advisory GHSA-3cp7-c6q2-94xr, pull request 20904, and release v0.5.10 contain the remediation that addresses the unsafe deserialization path in scheduler_client.py.

The associated EPSS score has remained flat at 0.0194 since disclosure, indicating no material rise in observed exploitation interest; the issue affects an LLM serving framework and is detailed further in reporting from Orca Security.

OWASP Top 10 for Web (2025)

EU & UK References

Vulnerability details

SGLang's multimodal generation module is vulnerable to unauthenticated remote code execution through the ZMQ broker, which deserializes untrusted data using pickle.loads() without authentication.

CWE(s)

Related Threats

MITRE ATT&CK Enterprise TechniquesAI

T1190 Exploit Public-Facing Application Initial Access
Adversaries may attempt to exploit a weakness in an Internet-facing host or system to initially access a network.
Why these techniques?

Unauthenticated remote code execution via deserialization in a network-accessible ZMQ broker in SGLang directly enables exploitation of a public-facing application.

Confidence: HIGH · MITRE ATT&CK Enterprise v19.0

CVEs Like This One

CVE-2026-3060Same product: Lmsys Sglang
CVE-2026-7301Same product: Lmsys Sglang
CVE-2026-7304Same product: Lmsys Sglang
CVE-2026-7302Same product: Lmsys Sglang
CVE-2026-5760Same product: Lmsys Sglang
CVE-2025-62368Shared CWE-502
CVE-2025-68903Shared CWE-502
CVE-2025-67911Shared CWE-502
CVE-2025-54014Shared CWE-502
CVE-2026-22505Shared CWE-502

Affected Assets

lmsys
sglang
0.5.5 — 0.5.9

Mitigating Controls

Mitigating Controls (NIST 800-53 r5) AI

prevent

Directly blocks unauthenticated clients from reaching the ZMQ broker and invoking pickle.loads().

prevent

Requires validation or rejection of untrusted serialized data before it is passed to pickle.loads().

prevent

Restricts network connectivity to the ZMQ broker so that only authorized endpoints can submit payloads.

References