Here's another way of spinning the same data: There have been more deaths caused by terrorism in the US in last 15 years than in the prior 135 years.
Furthermore, I don't think your comparisons to deaths from car accidents or cancer are completely valid. Those are both phenomena for which induction works well (what has happened previously is a good predictor of what will happen in the future) because the events come from more or less stationary distributions. As illustrated by September 11th attacks, the number of terrorist deaths can double in a single day. There's no reason that this can't happen again with an event with >5k deaths in a single day.
For those who have read anything by Nicholas Nassim Taleb: Nicholas Nassim Taleb might argue that terrorist attacks are black swan events (impossible to predict, potentially very large impact) and that the comparision of terrorist deaths (a distribution from Extremistan) to motor vehicle deaths is not very valuable.
On the other hand, I don't believe the level of NSA surveillance is appropriate. I'm just trying to illustrate the amount of effort one puts into thwarting something shouldn't just depend on its previous costs, but also on the probability distribution of the event (and your uncertainty in that).
tl;dr: Terrorism is more like pg's view of a startup in that it has potential for explosive growth, while motor accident deaths are more like a well established medium sized business.