There is a design choice at the heart of every assistant, and it is usually made against you. The easy path is to always sound sure. A confident answer feels helpful, tests well, and keeps people coming back. The trouble is that certainty and correctness are not the same thing, and in our field the gap between them is where people get hurt.

We took the other path deliberately. When our assistant is not confident, it says so, and it tells you why - the question is ambiguous, the situation is one where the safe move is to check with a human, or the honest answer is ‘it depends and here is what it depends on'. That is less slick. It is also the truth, and in security the truth is the only thing worth optimising for.

Consider the alternative. Someone asks whether a message is a scam, and a breezy ‘no, that looks fine' is wrong. They act on it. A hedged ‘I am not certain, but here are the two things that would tell you' costs them a minute and saves them a loss. Confident nonsense is worse than no answer, because it removes the caution that would otherwise have protected them.

We would rather you trusted the assistant less often and more accurately. When it is sure, you can lean on it. When it flags doubt, that flag is information, not a failure - it is the moment to slow down. We think that is what an honest tool owes the person using it, and we are not willing to trade it for the appearance of being clever.