Misaligned agents are just one layer of the AI security challenge that startup Witness AI is trying to solve. It detects employee use of unapproved tools, blocks attacks, and ensures compliance.