Robots can crack safes faster than humans—and differently:
So Seidle started looking for shortcuts. First he found that, like many safes, his SentrySafe had some tolerance for error. If the combination includes a 12, for instance, 11 or 13 would work, too. That simple convenience measure meant his bot could try every third number instead of every single number, immediately paring down the total test time to just over four days. Seidle also realized that the bot didn’t actually need to return the dial to its original position before trying every combination. By making attempts in a certain careful order, it could keep two of the three rotors in place, while trying new numbers on just the last, vastly cutting the time to try new combinations to a maximum of four seconds per try. That reduced the maximum bruteforcing time to about one day and 16 hours, or under a day on average.
But Seidle found one more clever trick, this time taking advantage of a design quirk in the safe intended to prevent traditional safecracking. Because the safe has a rod that slips into slots in the three rotors when they’re aligned to the combination’s numbers, a human safecracker can apply light pressure to the safe’s handle, turn its dial, and listen or feel for the moment when that rod slips into those slots. To block that technique, the third rotor of Seidle’s SentrySafe is indented with twelve notches that catch the rod if someone turns the dial while pulling the handle.
Seidle took apart the safe he and his wife had owned for years, and measured those twelve notches. To his surprise, he discovered the one that contained the slot for the correct combination was about a hundredth of an inch narrower than the other eleven. That’s not a difference any human can feel or listen for, but his robot can easily detect it with a few automated measurements that take seconds. That discovery defeated an entire rotor’s worth of combinations, dividing the possible solutions by a factor of 33, and reducing the total cracking time to the robot’s current hour-and-13 minute max.
We’re going to have to start thinking about robot adversaries as we design our security systems.