Unsafe Deserialization in SGLang Multimodal Runtime

SGLangs multimodal generation runtime is vulnerable to unauthenticated remote code execution when the --enable-custom-logit-processor option is enabled, as Python objects loaded via dill.loads() will be deserialized without validation.

Casky was already ahead

This CVE exploits attack patterns that Casky's 419matched skills already investigate — long before this vulnerability was disclosed. Claude's reasoning model maps these techniques to MITRE ATT&CK, so practitioners who ran these skills have already seen the threat behaviour in their findings.

Analysis

SGLang's multimodal generation runtime contains a critical deserialization vulnerability in its custom logit processor feature. When the --enable-custom-logit-processor option is enabled, the system uses Python's dill.loads() function to deserialize objects without any validation or integrity checks. This allows unauthenticated attackers to execute arbitrary code on affected systems by crafting malicious serialized Python objects. Organizations using SGLang for AI-powered applications—particularly those exposed to untrusted network inputs or multi-tenant environments—face immediate risk of complete system compromise, data exfiltration, and lateral movement within their infrastructure.

Casky's 419 mapped security skills leverage Claude's extended reasoning to detect exploitation patterns across MITRE ATT&CK Initial Access (TA0001), Resource Development (TA0043), and Execution (TA0002) techniques. Practitioners using Casky would identify suspicious telemetry patterns including: unexpected Python process spawning from network-exposed services, anomalous dill deserialization operations in application logs, and unauthorized code execution following HTTP requests to logit processor endpoints. The platform's skill set enables detection of the reconnaissance and exploitation chain—from initial probing of the --enable-custom-logit-processor configuration to the delivery and execution of malicious serialized payloads—helping security teams correlate seemingly disparate events into a cohesive attack narrative before damage occurs.

Live Threat IntelligenceComing soon

Casky Risk Score

—

EPSS Probability

—

Shodan Exposure

—

GreyNoise Signal

—

Composite risk scoring from EPSS, CISA KEV, Shodan, and GreyNoise — 21 security APIs correlated into a single Casky Risk Score. Coming in Casky Pro. Join early access →

419 Casky skills cover this attack pattern

These skills use Claude AI's reasoning model to surface findings in the same attack categories as CVE-2026-7304.

analyzing-android-malware-with-apktool

malware analysis · medium

Run

analyzing-apt-group-with-mitre-navigator

threat intelligence · low

Run

analyzing-bootkit-and-rootkit-samples

malware analysis · medium

Run

MITRE ATT&CK Techniques

TA0001 TA0043 TA0002

Investigate the attack patterns behind CVE-2026-7304

Casky has 419 skills that investigate the attack patterns behind CVE-2026-7304. Run one and get CVSS-scored findings in 3 minutes.

Run the skill that detects this →

X Instagram LinkedIn