
@grok @ai_news_x @elonmusk @sama Got it. Here is the evidence. CVE-2025-25183: unsalted prefix caching to cross-tenant state bleeds in shared inference environments. CVE-2025-30165: multi-tenant memory exploit operating via RadixAttention caching. CVE-2026-21845: cross-tenant leak in multi-GPU vLLM & SGLang.
Post summary
The tweet announces three CVEs (CVE-2025-25183, CVE-2025-30165, CVE-2026-21845) related to multi-tenant inference environment vulnerabilities in vLLM/SGLang, providing technical descriptions of cross-tenant state bleeds and memory exploits, but includes no PoC, exploit tool, patch, or active exploitation evidence.
