Machine learning (ML)-based attacks have shown that even the most advanced logic locking methods leave structural traces in circuits due to optimizations of security-agnostic synthesis tools. However, these attacks mainly focus on purely combinational logic locking techniques, like XOR/XNOR-based or MUX-based approaches. To mitigate structural leakages caused by synthesis tools, a novel flip-flop-based approach named Flip-Lock has been introduced. By incorporating sequential elements into locking substructures, this flip-flop-based logic locking approach prevents synthesis optimizations, effectively neutralizing all state-of-the-art structural attacks. This paper demonstrates that while flip-flop-based logic locking effectively mitigates the structural leakages left by synthesis tools; it may introduce flip-flop-relevant structural traces. These traces can reveal distinctions between the added locking flip-flops and the original flip-flops of the circuit. By reframing the problem as flip-flop classification rather than key extraction, we introduce Flip-Break, an Oracle-less Graph Neural Network (GNN)-based attack designed to target flip-flop-based logic locking. Flip-Break reveals that the added secure flip-flops differ from the original flip-flops in their local sub-structures, as they are randomly inserted into the circuit. We show that Flip-Break gains an average accuracy of in breaking add-on flip-flop-based secure layers for various benchmark circuits.