The problem with finding the longest path is that you have to explore the entire set of possible paths. A* works by finding an optimal list of steps that reach a known goal, and "get as far away as possible from policemen" isn't properly known in advance.
You could set a "comfortable" range where you consider yourself "safe" from policemen. If your policemen's A* heuristic function is "Distance to target" (where target is the thief), then your thief's heuristic would be "SafeRange - Distance to target" (where target is the policeman).
Now, since you have several policemen, you'd have to get distances to all of them, and possibly use the closest (most dangerous) one in the heuristic, or maybe add all of them together, or some smarter combination.
These are all ideas you can give a thought or try out. The best solution really depends on the way the maze is designed, and what other goals the thief has.