EIP-3074: AUTH and AUTHCALL opcodes

On balance, what msg.sender/tx.origin/gaspayer scheme do you (or anyone reading this) think makes the most sense in a post-merge, 3074-adopted world?

I’d love to see tx.origin aliased to msg.sender, personally. Elegant. Simple. Not sure how realistic.

Seems like there are two implementations of this so far: quilt go-ethereum (adeitrichs) and ethereumjs (brouwer)

Are we aware of any others? Is there some common location (aside from this thread) for implementation writers to report their work on 3074?

Can my suggestion for setting tx.origin to the account being remote-controlled, not the gas payer, be added to the EIP?

I’m worried that adding an even more controversial change to the EIP will make it even less likely to be adopted.

Once the merge is complete, I think we can reopen the discussion with core devs and see where we stand?

But it sounds like such a change is not controversial; more that nobody really seems to have an opinion one way or the other.

In this way, it’s more controversial not to make the change, isn’t it? Because I do have a very strong opinion on it.

Once the merge is complete, I think we can reopen the discussion with core devs and see where we stand?

Yes, I plan to voice my concern with the core devs group whenever it’s appropriate. Thanks for the reply.

From a perspective of secure contract design, I think having a shortcut for “remove authorisation” (= clear authorisation context) would be a nice addition, even at the expense of an extra opcode.

This can be current accomplished by sending “rubbish” to the AUTH instruction. Would auth(0, 0, 0) do the trick? The spec seems lenient at invalid input lengths.

Since the data in memory is not “versioned” or distinguished in any manner I do not think this achieves the goal of extensibility. I did found the old stack version much more “user friendly” too.

Yes, I believe so:

If the signature is instead invalid or the signer address does not equal authority, authorized is reset to an unset value.

In a future extension of AUTH, the interpretation of the memory range would be defined by authority (from the stack.) If there is code at that address, that code could parse the signature and decide to authorize or not. This is somewhat similar to EIP-1271’s isValidSignature. If no code exists at authority, the behaviour would be the same as 3074 currently specifies.

Me too, honestly.

I’d second this :+1:

@SamWilsn and authors, great EIP. I wonder if this could make Shanghai

1 Like

Now that the merge is complete, is now the right time to discuss how to handle tx.origin? The range of possibilities:

  1. Alias to msg.sender
  2. Set as “first actor”
  3. Leave as “gas payer”

I’d love to see option 1 (what are the implications of going this route?), but between choices 2 and 3, I think you are correct that option 2 gets us closer to true account abstraction and is, also, what I’m advocating for.

(I’ve joined the wallet dev discord and we can discuss there, if you like.)

Just to restate my opinion which I suspect is somewhere above: tx.origin referring to the signer/account that initiated the transaction makes the most semantic sense with “transaction origin”. However, this has caused much trouble because people use it to block “contracts” from interacting with dapps which includes contract wallets, and eventually could include account abstraction. I would feel no remorse breaking apps that have depended on this behavior as we have been advising people against this for a long time.

Respectfully but emphatically disagree.

3074 introduces something “new” - a rearward step in the transaction chain - and then moves something “old” (tx.origin) to this new first frame that never existed before. That really flies in the face of immutability and should be avoided at all costs, IMO.

I agree that we shouldn’t feel much remorse for those who were advised against tx.origin and then proceeded to use it anyway. But that’s not my case. The opposite, in fact. In 2015, I asked Chris Reitweissner about refunding the transaction sender and was told about tx.origin.

While he didn’t outright advise using tx.origin, you can see that it was floated as valid when I was previously on the msg.sender bandwagon. (The first two versions of Etheria correctly use msg.sender to determine token ownership. The last two use tx.origin, unfortunately, based on this gitter exchange.)

The danger of tx.origin wasn’t widely known until 9 months later in mid-2016 as shown here:

Indeed, leaving the EIP-3074 spec with tx.origin as the gaspayer would open a potential vulnerability: If a tile holder sponsors a transaction, any other frame in the transaction chain would have the ability to hijack these tokens. (Of course, we can advise owners to never use such an account to sponsor transactions, but I shouldn’t have to teach this nor can I ever guarantee reaching owners with this warning/education.)

If we are otherwise fine with breaking tx.origin because “people should have known better since mid-2016”, then we can both get what we want by simply aliasing tx.origin to msg.sender and being done with it. Solves my problem and eliminates tx.origin issues forever. (That said, I don’t have the expertise to know how impactful this would be throughout the entire ecosystem.)

Failing this abrupt aliasing, I strongly insist that “tx.origin” means, in spirit, “the first actor” and should remain that way for immutability reasons. This new initial frame created by 3074 should then be called “tx.gaspayer” to appropriately add a new label to a new frame and leave the old label with the old frame. Doesn’t this move us to a more complete “account abstraction” anyway?

Before thinking of altering the behaviour of tx.origin, please think about its current nice side-effect: it allows creating view-only transactions.
Assume that I want to simulate a specific code portion (either just to find if it runs at all, or some side-effects, such as creating contracts and query some data from them) - but WITHOUT really letting the user run this code on-chain.
The simple way we can use right now is to add require(tx.origin == address(0)) to the code (and document that this method has to be called only with “eth_call({from:'0x0' })

Do you know of any other way of forcing a code to run in view-mode only?

You’d need to somehow generate a valid signature from the zero address for that to be a problem. If someone can do that, we have bigger issues :wink:

1 Like

Creating a fake tx using an extremely high gas price or gas limit would most likely work (I do not think eth_call checks those things). And then thus check require(gasleft() > HIGH_VALUE) or require(tx.gasprice > HIGH_VALUE).

I have a quick question about AUTHCALL not resetting AUTH. From a security point of view, I would argue that it should reset AUTH; the reason is that now it opens ways for “attacks” similar as to re-entrancy attacks. (First AUTHCALL into some contract, do some nasty stuff over there to trick the calling contract to invoke another AUTHCALL when it returns from the authcall frame later…?)

You should only be signing AUTH messages to contracts (invokers) you explicitly trust to have 100% control of all of your funds. The idea here is that you can delegate control of your EOA to a contract, and you shouldn’t be AUTHing some random contract. Wallets should be incredibly restrictive about allowing this opcode, likely only allowing it for their own contract wallet or community trusted ones.

I think you mean to say you should only signing AUTH-messages to contracts (invokers) which you trust 100%. Once the AUTHCALL happens, it’s just the standard security model of interacting with smart contracts - e.g. they have control over only what you explicitly give them.

1 Like

Yes, fixed! More words to satisfy spam bot.

1 Like

I was always too skeptical about this EIP due it is introducing ability to steal all user tokens with 1 transaction, not only ETH as previously.

But today I realized that this EIP is effectively generalizing signature based calls as EIP-2612 (aka Permit) and others.

This EIP is also almost resolving EOA non-programmability.

1 Like