How would you know whether the agent is being used by a legitimate app or a malicious app racing with a legitimate app to steal access?

At least you only would leak a single access, and you would have a higher chance of noticing, but I can also see that if the hijack was done intermittently you might write it off as a glitch...