Training a 22MB prompt injection classifier that runs inline in a Lambda

How do you think catching slow and low attack