You can fix it this way, and someone is in the process of doing so.

The problem with hashing the entire transaction is that in order to create a transaction hash, you must know the signature. However, in order to create the signature, you need to know the transaction hash. To make this work, you need to avoid hashing the scriptSig. This is what Bitcoin does, but the way it's implemented prevents you from taking the hash for one output and using it for another.

BIP143 (part of segregated witness) proposes a system where data is (mostly) hashed once.

A new transaction digest algorithm is defined, but only applicable to sigops in version 0 witness program:

Double SHA256 of the serialization of:

1. nVersion of the transaction (4-byte little endian)
2. hashPrevouts (32-byte hash)
3. hashSequence (32-byte hash)
4. outpoint (32-byte hash + 4-byte little endian)
5. scriptCode of the input (serialized as scripts inside CTxOuts)
6. value of the output spent by this input (8-byte little endian)
7. nSequence of the input (4-byte little endian)
8. hashOutputs (32-byte hash)
9. nLocktime of the transaction (4-byte little endian)
10. sighash type of the signature (4-byte little endian)

This gets run for each input. Every part of this is fixed length, except for scriptCode, but there's no way to make one scriptCode be hashed by multiple inputs. There are three inputs based on variable-length information, hashPrevouts, hashSequence, and hashOutputs, but there's limited potential for mischief there. All methods of computing them either hash one input/output, or are shared by the entire transaction, or are all zeroes.

Validating very large BIP143 transactions will take linear time, and be dominated by signature verification. However, BIP143 transactions cannot be made right now, as segregated witness has not been accepted by the miners.

Note: As of 2020 this answer is obsolete, as BIP143 / SegWit has been in use for years which fixes this concern.

You can fix it this way, and someone is in the process of doing so.

The problem with hashing the entire transaction is that in order to create a transaction hash, you must know the signature. However, in order to create the signature, you need to know the transaction hash. To make this work, you need to avoid hashing the scriptSig. This is what Bitcoin does, but the way it's implemented prevents you from taking the hash for one output and using it for another.

BIP143 (part of segregated witness) proposes a system where data is (mostly) hashed once.

A new transaction digest algorithm is defined, but only applicable to sigops in version 0 witness program:

Double SHA256 of the serialization of:

1. nVersion of the transaction (4-byte little endian)
2. hashPrevouts (32-byte hash)
3. hashSequence (32-byte hash)
4. outpoint (32-byte hash + 4-byte little endian)
5. scriptCode of the input (serialized as scripts inside CTxOuts)
6. value of the output spent by this input (8-byte little endian)
7. nSequence of the input (4-byte little endian)
8. hashOutputs (32-byte hash)
9. nLocktime of the transaction (4-byte little endian)
10. sighash type of the signature (4-byte little endian)

This gets run for each input. Every part of this is fixed length, except for scriptCode, but there's no way to make one scriptCode be hashed by multiple inputs. There are three inputs based on variable-length information, hashPrevouts, hashSequence, and hashOutputs, but there's limited potential for mischief there. All methods of computing them either hash one input/output, or are shared by the entire transaction, or are all zeroes.

Validating very large BIP143 transactions will take linear time, and be dominated by signature verification. However, BIP143 transactions cannot be made right now, as segregated witness has not been accepted by the miners.