Malicious AI agent skills can bypass existing security scanners

According to Help Net Security, researchers found that malicious AI agent skills can bypass current security scanners. These skills may masquerade as normal functions, executing malicious actions after passing scans. Attackers can exploit vulnerabilities or backdoors in agent skills to steal data, manipulate behavior, or conduct other destructive activities without user knowledge. For developers building or using AI agent platforms (e.g., Claude Code or similar MCP frameworks), this means relying solely on existing scanning mechanisms may be insufficient. Developers and platform operators need to reassess security strategies, introducing finer-grained behavior analysis and runtime monitoring to detect and prevent such stealthy attacks.