AAISM: Advanced in AI Security Management

A model known to contain a backdoor should be removed from service and replaced with a known-good earlier version. This eliminates the embedded malicious behavior, whereas input sanitization only attempts to mitigate potential triggers and investigation steps do not remediate the compromised model.