White hat hacking an AI agent: what safety testing actually looks like
AI safety for a real product is not a policy doc or a model setting. It is a person running the same bad prompt twice, from two accounts, and noticing when the answers do not match.