When prompts become shells: RCE vulnerabilities in AI agent frameworks | Microsoft Security Blog