Without a sharp constraint on the maximum blocksize there is currently _no_ rational reason to believe that Bitcoin would be secure at all once the subsidy goes down.
This is a very interesting game-theory question.
I have done some preliminary analysis of the problem and have found that gmaxwell may not be correct on this assertion. (I used to support the razing of the max block size limit; however I now reject it on moral ground. The same moral grounds that I would reject any change of the number of coins; that it would make any past uses of the protocol under a false pretence)
Now I will try and explain why a bitcoin-like protocol could be secure without the max-block-size limit.
The core issue is the mixing up cost-to-attack with active hash rate, when in the long run they separate to be quite independent qualities. While the vast majority of the miners income is from mining blocks for new bitcoin; there happens to be a very strong causation, this causation doesnt need to hold.
The first thing that we can define is the damage cost of double spends this cost can be modelled defined by the equation:
cost = time x value
time: real number between 0 and 1
A double spend a long time after a transaction, for a large amount is a very costly (up to the full cost of the tx).
In the free-market, in the long term, the market will sustain any fraud rate that is less than the cost of reducing it. (aka, it is cheaper to buy insurance than to increase the difficulty of an attack).
I see no reason why a bitcoin-like protocol wouldn't be subjected to the same principles: The network will spend the absolute minimum on maintaining its security. (Either via hashing or via insurance).
So what is the cheapest form of network security? Dynamic response to threats.
Bitcoin miners incur virtually no cost; unless they are actively mining. Bitcoin insurance companies could amass huge collections of bitcoin miners and turn them on when it is cheaper for them to out-mine the double-spend than pay the insurance out.
The bitcoin network will look quite easy to attack; well until you try and attack it.
This will raise the COST TO ATTACK (that is a constant); while the COST TO DEFEND is at a minimum; only the minimum number of miners are turned on to defend the chain when it is attacked.
Otherwise a background mining operation will be run by bitcoin companies for general network health.