Meta Muse macOS 本地零日漏洞 PoC 公开

Dark Web Intelligence @DailyDarkWeb原推发布于 2026.09.22

在 X 上打开原推

🚨 META MUSE ZERO-DAY LETS LOCAL MALWARE HIJACK THE AI AGENT’S TRUST AND ACCESS macOS security researcher Patrick Wardle has released a public PoC for not-a-mused, a local zero-day affecting Meta’s Muse AI assistant for macOS. • Muse exposes an undocumented setting, endo_voyager_dictation_endpoint • An unprivileged local process can modify the endpoint without administrator privileges • Redirecting it can send dictated prompts to an attacker-controlled server • Wardle’s PoC shows potential capture of prompts, prompt injection, theft of Muse authentication material and abuse of access already granted to the agent • The attack requires existing local code execution — this is NOT a remote zero-click compromise • The security impact is amplified because an AI agent may hold substantially broader permissions than the malware that initially lands on the Mac • The researcher’s PoC implements a subset of more than 50 commands exposed by Muse • No confirmed in-the-wild exploitation has been identified • No public CVE or confirmed patched version was identified at the time of disclosure ⚠️ Analyst Note: This is a strong example of AI agents becoming privilege amplifiers. Traditional macOS controls may prevent ordinary malware from directly accessing sensitive resources, but if that malware can hijack a highly privileged agent already trusted by the user, the agent itself can become the attacker’s bridge into files, services and other authorized capabilities. Original researcher PoC: https://github.com/pwardle/not-a-mused #Meta #Muse #AISecurity #AgenticAI #macOS #ZeroDay #ThreatIntel #CyberSecurity #DDW #DarkWeb