The `signer` binary contains a ruleset engine, implemented with [OttoVM](https://github.com/robertkrimen/otto)
It enables usecases like the following:
* I want to auto-approve transactions with contract `CasinoDapp`, with up to `0.05 ether` in value to maximum `1 ether` per 24h period
* I want to auto-approve transaction to contract `EthAlarmClock` with `data`=`0xdeadbeef`, if `value=0`, `gas < 44k` and `gasPrice < 40Gwei`
The two main features that are required for this to work well are;
1. Rule Implementation: how to create, manage and interpret rules in a flexible but secure manner
2. Credential managements and credentials; how to provide auto-unlock without exposing keys unnecessarily.
The section below deals with both of them
## Rule Implementation
A ruleset file is implemented as a `js` file. Under the hood, the ruleset-engine is a `SignerUI`, implementing the same methods as the `json-rpc` methods
if (req.metadata.scheme == "ipc"){ return "Approve"}
}
```
Whenever the external API is called (and the ruleset is enabled), the `signer` calls the UI, which is an instance of a ruleset-engine. The ruleset-engine
invokes the corresponding method. In doing so, there are three possible outcomes:
1. JS returns "Approve"
* Auto-approve request
2. JS returns "Reject"
* Auto-reject request
3. Error occurs, or something else is returned
* Pass on to `next` ui: the regular UI channel.
A more advanced example can be found below, "Example 1: ruleset for a rate-limited window", using `storage` to `Put` and `Get``string`s by key.
* At the time of writing, storage only exists as an ephemeral unencrypted implementation, to be used during testing.
### Things to note
The Otto vm has a few [caveats](https://github.com/robertkrimen/otto):
* "use strict" will parse, but does nothing.
* The regular expression engine (re2/regexp) is not fully compatible with the ECMA5 specification.
* Otto targets ES5. ES6 features (eg: Typed Arrays) are not supported.
Additionally, a few more have been added
* The rule execution cannot load external javascript files.
* The only preloaded library is [`bignumber.js`](https://github.com/MikeMcl/bignumber.js) version `2.0.3`. This one is fairly old, and is not aligned with the documentation at the github repository.
* Each invocation is made in a fresh virtual machine. This means that you cannot store data in global variables between invocations. This is a deliberate choice -- if you want to store data, use the disk-backed `storage`, since rules should not rely on ephemeral data.
* Javascript API parameters are _always_ an object. This is also a design choice, to ensure that parameters are accessed by _key_ and not by order. This is to prevent mistakes due to missing parameters or parameter changes.
* The JS engine has access to `storage` and `console`.
#### Security considerations
##### Security of ruleset
Some security precautions can be made, such as:
* Never load `ruleset.js` unless the file is `readonly` (`r-??-??-?`). If the user wishes to modify the ruleset, he must make it writeable and then set back to readonly.
* This is to prevent attacks where files are dropped on the users disk.
* Since we're going to have to have some form of secure storage (not defined in this section), we could also store the `sha3` of the `ruleset.js` file in there.
* If the user wishes to modify the ruleset, he'd then have to perform e.g. `signer --attest /path/to/ruleset --credential <creds>`
##### Security of implementation
The drawbacks of this very flexible solution is that the `signer` needs to contain a javascript engine. This is pretty simple to implement, since it's already
implemented for `geth`. There are no known security vulnerabilities in, nor have we had any security-problems with it so far.
The javascript engine would be an added attack surface; but if the validation of `rulesets` is made good (with hash-based attestation), the actual javascript cannot be considered
an attack surface -- if an attacker can control the ruleset, a much simpler attack would be to implement an "always-approve" rule instead of exploiting the js vm. The only benefit
to be gained from attacking the actual `signer` process from the `js` side would be if it could somehow extract cryptographic keys from memory.
##### Security in usability
Javascript is flexible, but also easy to get wrong, especially when users assume that `js` can handle large integers natively. Typical errors
include trying to multiply `gasCost` with `gas` without using `bigint`:s.
It's unclear whether any other DSL could be more secure; since there's always the possibility of erroneously implementing a rule.
## Credential management
The ability to auto-approve transaction means that the signer needs to have necessary credentials to decrypt keyfiles. These passwords are hereafter called `ksp` (keystore pass).
### Example implementation
Upon startup of the signer, the signer is given a switch: `--seed <path/to/masterseed>`
The `seed` contains a blob of bytes, which is the master seed for the `signer`.
The `signer` uses the `seed` to:
* Generate the `path` where the settings are stored.
* Generate the encryption password for `vault.dat`.
The `vault.dat` would be an encrypted container storing the following information:
*`ksp` entries
*`sha256` hash of `rules.js`
* Information about pair:ed callers (not yet specified)
### Security considerations
This would leave it up to the user to ensure that the `path/to/masterseed` is handled in a secure way. It's difficult to get around this, although one could
imagine leveraging OS-level keychains where supported. The setup is however in general similar to how ssh-keys are stored in `.ssh/`.
# Implementation status
This is now implemented (with ephemeral non-encrypted storage for now, so not yet enabled).