CVE-2026-3059Disclosure(lmsys / sglang)

MEDIUMCVSS 9.8 · CRITICAL

Exploitation ongoing with high activity in latest observed window (1 mentions)

Immediate actions

Patch lmsys sglang systems immediately
Assume compromise if assets are exposed

Recommended action window: Immediate (within 24h)

NVD description

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.

lmsys CVEs Back to dashboard

4.3/ 10 priority

Sources & remediation

Exploit / PoC references

https://orca.security/resources/blog/sglang-llm-framework-rce-vulnerabilities/

Other references

Weakness type (CWE)

CWE-502

Priority

MEDIUM

Exploitation

ACTIVE

PoC

YES

Patch

AVAILABLE

Momentum

STABLE

Are you affected?

If you run products in this scope, you should treat this CVE as relevant to your environment.

sglang

Threat summary

Active exploitation appears in 1 classified signals
Patch or workaround signal is available
9 mentions across 8 observed days
Momentum state: stable

What's happening

Active exploitation reported across 1 signal
Patch or workaround mentioned in 1 signal
Technical details provided in 8 signals
Disclosure: 5 classified signals
General: 2 classified signals
Peaked 4d ago at 2 mentions (2026-03-17); latest day: 1
9 total mentions across 8 days

Affected systems

Vendors

lmsys

Products

sglang

Deep dive

Activity timeline9 mentions / 8d

Signal classification4 categories

Disclosure

555.6%

General

222.2%

Patch

111.1%

Active Exploitation

111.1%

Referenced assets7 URLs

By indicator

Classification over time

Date	Total	Labels

2026-03-12	1	Disclosure1
2026-03-13	1	Disclosure1
2026-03-16	1	Disclosure1
2026-03-17	2	Disclosure2
2026-04-09	1	Patch1
2026-04-16	1	General1
2026-04-17	1	General1
2026-04-21	1	Active Exploitation1

Full discourse9 posts

Gray Hats@the_yellow_fall206d
Disclosure
Critical 9.8 CVSS unpatched RCE flaws in the SGLang AI framework (CVE-2026-3059 & CVE-2026-3060) expose servers via unsafe Python pickle deserialization. #SGLang #AISecurity #PickleDeserialization #CVE #LLMSecurity #CyberSecurity #InfoSec #RCE #ZeroDay https://securityonline.info/poisoned-pickle-critical-unpatched-rce-flaws-sglang-ai-infrastructure/ https://t.co/ZAln4BRyXc
Post summary
The tweet announces two unpatched RCE vulnerabilities (CVE‑2026‑3059 & CVE‑2026‑3060) in the SGLang AI framework, highlighting high severity (CVSS 9.8) and unsafe pickle deserialization, but provides neither PoC, exploit code, active exploitation claims nor patch information.
02041456
10.7K followersView on X
O Pantera Negra · IA & Produto 🤖@opanteranegra77175d
General
SGLang RadixAttention vence benchmarks (29% mais rápido), mas CVE-2026-3059 é RCE unauthenticada na multimodal. Testei: infrastructure attack surface > performance gains. Otimização sem segurança = piscina com hiena. #IA #PromptEngineering
Post summary
The post confirms CVE-2026-3059 as an unauthenticated RCE, but does not provide a PoC, exploit, active exploitation evidence, or patch information.
3000037
232 followersView on X
Inferlume@inferlume_hq170d
Active Exploitation
The other two (CVE-2026-3059, 3060): pickle deserialization over the network. Pre-auth. No credentials needed. Just send a crafted payload. No patch for these either. SentinelLabs says: unexpected outbound connections from SGLang workers = you've already been hit.
Post summary
CVE-2026-3059 and 3060 allow pre‑auth pickle deserialization over the network; SentinelLabs reports remote exploitation via unexpected outbound connections from SGLang workers, with no patches or workarounds available.
1000027
1 followersView on X
CVETrends@CVEShield174d
General
Top 5 Trending CVEs: 1 - CVE-2018-17144 2 - CVE-2026-34621 3 - CVE-2010-5139 4 - CVE-2026-32201 5 - CVE-2026-3059 #cve #cvetrends #cveshield #cybersecurity https://www.cveshield.com/dashboard
Post summary
The post lists five CVEs as trending without providing any further details, claims, or technical information.
00010148
1.7K followersView on X
PulsePatch.io@pulsepatchio183d
Patch
`SGLang`'s multimodal generation module is vulnerable to unauthenticated RCE via the ZMQ broker (CVE-2026-3059). Restrict network access to the ZMQ port and monitor for a fix. #SGLang #RCE #infosec https://www.pulsepatch.io/posts/cve-2026-3059-sglang-rce-zmq-broker
Post summary
SGLang’s multimodal generation module has an unauthenticated RCE (CVE‑2026‑3059); restricting access to the ZMQ port is advised while vendors address the issue.
0000042
11 followersView on X
SH TC@shtc_social205d
Disclosure
🚨 Amazon Bedrock, LangSmith ve SGLang'de Kritik AI Güvenlik Açıkları! DNS sorguları ile veri sızdırma ve RCE saldırıları mümkün. CVE-2026-3059/3060 (CVSS 9.8) kritik! #SiberGüvenlik #AI #AmazonBedrock #LangSmith #SGLang #CVE 🔗 https://sh.tc/amazon-bedrock-langsmith-sglang-ai-guvenlik-aciklari
Post summary
The tweet discloses critical CVE-2026-3059/3060 affecting Amazon Bedrock, LangSmith, and SGLang, noting potential DNS-based data exfiltration and remote code execution risks with a CVSS score of 9.8; no PoC, exploit tool, patch, or active exploitation evidence is provided.
0000066
109 followersView on X
SH TC@shtc_social205d
Disclosure
🚨 AI Platformlarında Kritik Güvenlik Açıkları! Amazon Bedrock, LangSmith ve SGLang'de keşfedilen zafiyetler veri sızıntısı ve uzaktan kod çalıştırma riski oluşturuyor. CVE-2026-3059 CVSS 9.8! #SiberGüvenlik #AI #CVE #CyberSecurity 🔗 https://sh.tc/ai-platformlari-kritik-guvenlik-aciklari-bedrock-langsmith-sglang
Post summary
The post announces newly discovered CVE-2026-3059 in AI platforms with a high CVSS score, highlighting potential data leakage and remote code execution risks, but provides no PoC, exploit code, or patch information.
0000067
109 followersView on X
CVE@CVEnew209d
Disclosure
CVE-2026-3059 SGLang's multimodal generation module is vulnerable to unauthenticated remote code execution through the ZMQ broker, which deserializes untrusted data using pickle.load… https://www.cve.org/CVERecord?id=CVE-2026-3059
Post summary
CVE-2026-3059 reveals an unauthenticated remote code execution vulnerability in SGLang's multimodal generation module, stemming from unsafe pickle deserialization in the ZMQ broker.
00000152
56.7K followersView on X
The Hacker Wire@TheHackerWire210d
Disclosure
🔴 CVE-2026-3059 - Critical SGLang's multimodal generation module is vulnerable to unauthenticated remote code execution through the ZMQ broker, which deserializes untrusted data using pickle.loads() without authenti... https://www.thehackerwire.com/vulnerability/CVE-2026-3059/ https://t.co/HsQEWj8V2B
Post summary
SGLang’s multimodal generation module is exposed to a critical unauthenticated remote code execution vulnerability caused by deserializing untrusted data with pickle.loads() via its ZMQ broker.
0000037
134 followersView on X

CPE platform detail1 entries

Part:

Vendor:

Product:

Version:

Target SW:

Target HW:

1 of 1 entries

Part	Vendor	Product	Version	Target SW	Target HW
App	lmsys	sglang	-	-	-

Explore more

Back to dashboard lmsys CVEs sglang vulnerabilities

What researchers are saying

CVE-2026-3059

O Pantera Negra · IA & Produto 🤖[verified]@opanteranegra77General

The post confirms CVE-2026-3059 as an unauthenticated RCE, but does not provide a PoC, exploit, active exploitation evidence, or patch information.

Inferlume[verified]@inferlume_hqActive Exploitation

CVE-2026-3059 and 3060 allow pre‑auth pickle deserialization over the network; SentinelLabs reports remote exploitation via unexpected outbound connections from SGLang workers, with no patches or workarounds available.

SH TC[verified]@shtc_socialDisclosure

The tweet discloses critical CVE-2026-3059/3060 affecting Amazon Bedrock, LangSmith, and SGLang, noting potential DNS-based data exfiltration and remote code execution risks with a CVSS score of 9.8; no PoC, exploit tool, patch, or active exploitation evidence is provided.

SH TC[verified]@shtc_socialDisclosure

The post announces newly discovered CVE-2026-3059 in AI platforms with a high CVSS score, highlighting potential data leakage and remote code execution risks, but provides no PoC, exploit code, or patch information.

Gray Hats@the_yellow_fallDisclosure

The tweet announces two unpatched RCE vulnerabilities (CVE‑2026‑3059 & CVE‑2026‑3060) in the SGLang AI framework, highlighting high severity (CVSS 9.8) and unsafe pickle deserialization, but provides neither PoC, exploit code, active exploitation claims nor patch information.

CVETrends@CVEShieldGeneral

The post lists five CVEs as trending without providing any further details, claims, or technical information.

PulsePatch.io@pulsepatchioPatch

SGLang’s multimodal generation module has an unauthenticated RCE (CVE‑2026‑3059); restricting access to the ZMQ port is advised while vendors address the issue.

CVE@CVEnewDisclosure

CVE-2026-3059 reveals an unauthenticated remote code execution vulnerability in SGLang's multimodal generation module, stemming from unsafe pickle deserialization in the ZMQ broker.