Securing AI systems

An injection fools the model 20% of the time. An argument validator misses 10% of bad calls and human approval misses 5%. Assuming independence, what is the probability of harm per attempt?