Nerd sniping: tricking AI into overthinking its way into a wasteful loop
Ourtake
Teach the robot to think harder and someone will just hand it a puzzle it can never finish.
What happened
An OpenAI study tested o1-preview and o1-mini against manipulation attempts and found that more processing time generally improved resistance. The researchers also found new attacks, 'think less' and 'nerd sniping', which cut short or waste the models' reasoning time, and noted these are hard to spot.