Skip to content
CSA Labs Research·Indexed by this site · Originally published· Event →· Its place in the scan record →SelectedAttention44

Confidential AI Under Pressure: SGLang RCE and SEV-SNP Attack

Confidential AI Under Pressure: SGLang RCE and SEV-SNP Attacks

AI summary

NEAR AI disclosed that attackers exploited a remote code execution vulnerability in SGLang on its servers to execute code on inference servers and move laterally; AMD concurrently published two attacks, Mior and Skyfall, against SEV-SNP. NEAR AI said it exposed the names and emails of about 100 users, and the chat records and logs of 18 agents running on old non-TEE deployments; API keys of two NEAR AI Cloud customers had been written to logs and have been revoked; requests processed by the compromised server on November 8 included confidential-mode requests and may have been observed. AMD's Mior (CVE-2026-0462) lets a malicious host controlling the hypervisor and BIOS read and write guest memory and forge attestation reports; according to the researchers, all 1,000 attempts across two rounds of testing on an EPYC 9135 succeeded. Skyfall (CVE-2026-40671) leaks guest register state via a transient-execution side channel.

Why it was selected

The two types of incidents in the article come from the software and hardware sides respectively; the TCB check values and remediation timelines given can be used to verify the actual scope of protection of confidential inference services.

Read originallabs.cloudsecurityalliance.org