Cyber Risks · The Hacker News

Researchers Find Serious AI Bugs Exposing Meta, Nvidia, and Microsoft Inference Frameworks

Summary: Oligo Security researchers found critical remote code execution flaws in AI inference engines from Meta, Nvidia, Microsoft, vLLM, and SGLang, all traced to a pattern they call ShadowMQ: unsafe pickle deserialization over unauthenticated ZeroMQ TCP sockets. The root cause was a flaw in Meta's Llama framework (CVE-2024-50050, CVSS 6.3/9.3) patched last October, and Oligo found the same logic recurring elsewhere, in some cases through direct copy-paste, with SGLang's vulnerable file noting it was adapted from vLLM and Modular Max Server borrowing from both.

Key Takeaway: A single compromised inference node can let an attacker execute arbitrary code on the cluster, escalate privileges, steal models, or drop payloads like cryptocurrency miners - and the flaw spread because maintainers at different companies copied the same unsafe code.

Read the original at The Hacker News ↗

Topics
AI Security
Mentions
CVE-2024-50050
Related