Safety evaluation against malicious inputs