The CERT Coordination Center has disclosed VU#431821, a severe command injection vulnerability (CVE-2026-2256) in ModelScope's MS-Agent framework, allowing attackers to execute arbitrary OS commands via crafted prompt inputs. The flaw affects the Shell tool, where unsanitized user-derived content bypasses fragile denylist filters through techniques like encoding or obfuscation. Successful exploitation grants full remote code execution (RCE) privileges matching the agent's process, risking system compromise, data theft, or persistence. Discovered by researcher Itamar Yochpaz, no official patch exists yet, prompting urgent mitigation advice including sandboxing, least-privilege execution, and restricting agents to trusted sources. Proof-of-concept code is publicly available, heightening the risk for users processing unvalidated external content.