Start with a state machine for the drone and a finite‑state machine for the AI, then run a Monte Carlo sweep of the AI’s possible responses. Treat the AI as a Markov chain—each decision is a transition probability—and look for low‑entropy states where it always chooses the same action. Put the drone in those states, flip a coin on the next move, and you force the AI out of its comfort zone. Inject a tiny, unpredictable jitter into the drone’s speed or heading—just enough to shift the AI’s predicted path—and you’ll break the pattern. Finally, after a few runs, log the AI’s move history, find the recurring sequences, and program your drone to counter the first move of that sequence. Simple, but it turns the AI’s predictability into a weakness.