Sufficiently Decentralized OAuth Protocol for Both Security and Usability of Web3 | Dennis Won

[Music] hello everyone my name is Dennis from Magic labs and I'm here to talk about Newton key store um key store roll up with trustless verifiable onchain Automation in detail uh I'm going to go over the onchain and offchain hybrid infrastructure for agent-driven decentralized finance with very viable security and privacy at Magic we are very passionate about the convergence of AI and web 3 um web 3 and AI they individually respectively bring their own benefits and strengths web 3 reduces dependency on intermediaries uh prioritizes data sovereignty and privacy and provide censorship resistance AI on the other hand it automates tasks and optimize processes and is capable of solving complex problems at unprecedented scale we believe the future is an Asian deriven economy um system where autonomous AI programs like agents interact transact and execute tax independently on blockchain networks we believe um there will be defi Bots and decentralized marketplaces and governance those onchain processes to be automated and enhanced for user experience we have already seen many examples of AI coming on chain we have seen different like AI chatbots mainly for kind of social um social management Community Management slash data analytics however what we are at what what we believe um at Magic is we are at the verge of seeing agents actually taking actions on the behalf of users to solve the web three complexity problem and that's what we are trying to solve yeah so we believe uh this is the future of what's what we can achieve by combining crypto and AI web 3 will have enhanced defi with AI power analytics and fraud detection AI Power Smart contracts automating complex processes and personalized experience for users without compromising privacy not just chain abstraction but also app abstraction without web three complexity for AI trustless execution data Integrity tokenized incentives and interoperability to start off I would like to um you know start with an example the user says can you find the top 10 Traders on base in terms of p&l the last seven days the agent replies here are the top 10 Traders on base in terms of p&l the user says can new setup copy trade for me cross chain for these addresses with Max spent limit of 10,000 usdc in value until March 2nd 2025 and agent says sure I've set up copy trading function for your account with these permissions the problem is in order to really see the future of Agents taking action on the behalf of the user taking ACC accessing the user's funds the module execution Integrity becomes a question here how can we verify correct model execution how to preserve model privacy um without sacrificing privacy how to ensure against Mo U model tempering how can users and agents trust each other that's what we are trying to solve at Newton yeah so this is a diagram uh there's a trust Gap uh currently for client request um after making uh clients making the inference request yeah what is the model is it genuine execution any security concerns so this is what we provide um for users we provide chain abstracted 7 ERC 7579 modular smart contract Newton account it's gas optimized and fully audited and it's chain it comes with chain abstraction and the signer is secure with secure inclave uh trusted execution environment for users provide social login Pass Key good user experience for agents it provides remote attestation that can be used to uh perform verif to be verified um on chain underneath Newton key store is our um infr structure that in essence it is a cross-chain ZK key store rollup it operates ZK VM to provide programmable stateful and verifiable permissions also in order to enable verifiable automation for AI agents we have the uh unique model registry and verifiable infuence mean poool to provide censorship resist resistance trustless AI experience for users and agents so we Define this key store as the trust layer for agents with verifiable autom autonomous execution the verification um includes execution Integrity agent sovereignty proof-based decision making and sensorship resistance first of all I'll start with the AR of the first portion of the architect which is the key store rollup itself how we achieve the chain abstraction for our permission rollup for L uh ethereum and other l2s we operate like clients which uh are part of our Newton validator Network essentially they aggregate the key store contract state from all the connected rollups Contra and it uh manages these two muls for permissions and also inference separately it sequences the block which gets executed by the verifier node which operates ckvm with read optimized index Merk tree so that's where the permission State the program for programmable session key is being stored it also have model registry I'll talk about the model registry in a bit and in addition lastly it provides remote attestation verifying that the execution by the verifier node on zkm is correct the executive block gets returned back to the Newton volor Network which provides consensus and fast finality to be settled to ethereum and use ZK State proof to synchronize to all the other l2s and this is the second portion of verifiable automation which is the model registry part and how we achieve verifiable automation this is a combination of offchain and onchain infrastructure that uh that I mentioned earlier to begin with I know the words are the letters are quite small here um and I cannot zoom in to make it Lar larger but you know I'll I'll just speak it out so it starts from AI developers they use our SDK to um circu tize their model or if the model is running on te it also can use the remote testation of the te to get deploy as uh into a verifier contract on Newton for later verification of the execution so the model registry contains all the model details including input out uh output type the commitment which is uh what is the required commitments that the operators of the model should achieve and yeah lastly and other metadata so the our Network contain exact information that we can use uh in in addition to the Circuit that is deployed on our zkm to later used to verify the automation execution of the models of the inference and uh when at the time of the user the permission delegate the users want to make inference it submits inference request to a decentralized inference me poool that our Val ators operate and which orchestrates those inference requests exactly to the node oper uh model operators that register with a state essentially those model operators when they are registering they're making a commitment with the stake that can be slashed if their model execution fails to get verified on on our rollup so the again back uh back to the inference M Pool it gets orchestrated to a specific a model operators which provides the inference result back to the user um and if they want to they can also store their inference result to a ipfs or file coin or R weave like permanent Storage decentralized storage and in addition to the inference result they are responsible for providing a proof of valid execution of the model whether it's te remote attestation or the ZK proof that um that can only be verified by the verifier contract if they fail to uh get verified or if they fail to provide the proof they get slashed otherwise they um defeat they earn the fees that inference request um uh require and so by making this variable of inference this uh that's how they get authenticated with verifi uh with verific ZK verification to the users's smart accounts to trigger intents to execute intents on the behalf of the user also like many people would have uh um being are concerned about the privacy of the links between their wallets essentially like users manage multiple wallets their interaction with their agents all those interactions people don't want to make it public and store on chain so we also provide optional privacy for putting permissions on chain where on the client side they can execute in a private execution environment so which is essentially they updated the state uh on the client side off chain and then send a private transaction along with the proof of the execution correctness uh and broadcast it back to the node this is actually very similar to how Aztec provides private execution yeah so this is the last slide uh of my talk um with Newton account which is chain abstracted uh wallet a smart account for users also underneath with a key store roll up with verifiable automation we um provide for users chain abstraction but also app abstraction using AI cross chain verifile permissions transparent verified AI predictions for auto onchain automations and cryptography guarantee of faithful execution by the agents for agents we enable we're still running the AI models on offchain the inference we're not running AI models onchain however we are putting the verification circuit on chain and the permissions that agents are given with specific policies that are given by the user it is protect uh therefore the ml models they can protect their propriatary uh algorithms also thank uh thanks to the decentralized mle this can be also used for cross chain Agent form where this infrastructure can be used for multi-agent coordination and collaboration also with decentralized fee uh fee based like Marketplace for inference models are able to get secure payments from users decentralized yeah so that was the gist of the talk um it was um this is what we are trying to provide for both users and also agents trying to uh we believe uh these two together can be a fundamental infrastructure for crypto and AI convergence thank you any questions no