Submitted by oakcobalt, also found by muellerberndt and RadiantLabs
handle_l1_message can only be invoked by starknet os(@l1_handler) and is not a regular user invoked transaction from eth_send_raw_unsigned_tx flow.
kakarot::handle_l1_message -> [library:[handle_l1_message](https://github.com/kkrt-labs/kakarot/blob/7411a5520e8a00be6f5243a50c160e66ad285563/src/kakarot/library.cairo#L421) -> Interpreter::execute)
handle_l1_message hardcodes EVM gaslimit(2100000000) and gasprice(1) for every L1->L2 message regardless of the complexity of the actual l2 tx. In interpreter::execute, L1senders cached balance will be subtracted with the max_fee (2100000000 x 1) first before performing ETH transfer or running EVM.
Case: L1 sender performs minimal operations on L2:
For an L1 sender who only transfers some ETH to a L2 address or perform simple opcodes, the actual gas cost (required_gas) may be very close to the intrinsic gas cost(21000). This means the actual_fee l1sender is required to pay is around 21000 x 1, which is far less than the calculated max_fee 2100000000 x 1.
In this case, interpreter::execute will first subtract max_fee(2100000000) from L1senders cached balance (Account.set_balance(sender, &new_balance)). Note that this cached balance subtraction is done before ETH value transfer and run(evm), which means any subsequent logic will be using L1senders new_balance (e.g., X - 2100000000).
https://github.com/kkrt-labs/kakarot/blob/7411a5520e8a00be6f5243a50c160e66ad285563/src/kakarot/interpreter.cairo#L950-L951
Suppose L1senders current balance is (2100000000 + 21000). And L1sender wants to transfer 2100000000 to an L2 address. Suppose the required_gas is only 21000 so that L1senders current balance is sufficient.
The above shows that despite L1sender having enough balance to perform the L2 tx, L1senders tx is reverted with max fee 2100000000 charged. L1sender only has 21000 left after the tx and no transfer is done.
L1senders L2 tx can be unfairly reverted when they have enough balance to cover transfer value and the actual fee.
Because handle_l1_message is not invoked by RPC endpoint users, L1sender doesnt have a chance to specify max_fee value. In such cases, L1sender might only reasonably expect to have the EVM gas cost sufficient based on the complexity of the L2 tx, which can be much lower than max_fee. Although max_fee is marginal in ETH value, L1senders L2 tx can still fail due to the sequence of fee deduction in the cached state. The failure is not the L1senders fault.
Because handle_l1_message is different from eth_send_raw_unsigned_tx in its invocation and max_fee setting, consider allowing L1sender to input gas_limit from L1, and decode the user input gas_limit in handle_l1_message to compute max_fee. Or, refactor the control flow to skip max fee deduction when called from handle_l1_message.
ClementWalter (Kakarot) confirmed and commented:
Kakarot mitigated:
Status: Mitigation error. Full details in reports from RadiantLabs included in the Mitigation Review section below.
For this audit, 13 reports were submitted by wardens detailing low risk and non-critical issues. The report highlighted below by Radiant Labs was marked as best from the judge and is included for completeness. 
The following wardens also submitted reports: Rhaydden, 20centclub, Bauchibred, NexusAudits, Koolex, kirill_taran, 1AutumnLeaf777, Emmanuel, gumgumzum, oakcobalt, fyamf, ahmedaghadi. 
Please see the Mitigation Review section below for a complete list of all Low Risk and Non-Critical issues identified during the audit, and confirmed as mitigated. 
