Summary: AI agents, USD Coin (USDC), and zero knowledge proofs (ZKPs) are converging to bring about autonomous and trustless global payments. USDC’s frictionless global payment rails are a perfect match for autonomous AI agents that can transact business globally at anytime. ZKPs provide privacy-preserving verification of AI agent properties and models, and can trigger USDC payments based on verified off-chain compute. NovaNet ZKPs are well-suited for verifying AI agents and a wide variety of off-chain compute due their memory efficiency, portability, and ability to generate proofs locally on devices.
A new frontier is emerging at the intersection of AI, cryptocurrency, and privacy-preserving verifiable compute.
AI agents, USD Coin (USDC), and zero knowledge proofs (ZKPs) are converging to bring about autonomous and trustless global payments.
As noted by Circle CEO Jeremy Allaire, “USDC will almost certainly be the preferred currency of AI Agents.” We agree.
An artificial intelligence (AI) agent is software that is capable of making complex decisions and executing tasks autonomously. These tasks include anything from research and content development, completing workflows, negotiating transactions, and executing payments.
Enter USD Coin (USDC), a digital dollar that offers the perfect solution for AI-driven payments. As a stablecoin pegged to the US dollar, USDC combines the stability of fiat currency with the programmability and global reach of cryptocurrency.
At NovaNet, we built a ZKP system that brings privacy, verifiability, and security to AI agents that handle sensitive financial data and act autonomously.
There are several features of USDC that make it attractive to AI agents:
1. Ease of access: Unlike traditional banking systems with their complex APIs and regulatory hurdles, USDC can be easily integrated into AI systems.
2. Programmability: USDC, being a digital asset, can be programmed with smart contracts, allowing for complex, automated financial operations.
3. Speed and Global availability: USDC operates 24/7, processes transactions in seconds, and is not bound by geographical limitations. This is perfect for AI agents that operate across borders and time zones to settle payments instantly regardless of local banking hours or holidays.
4. Cost effectiveness: Lower transaction fees compared to traditional banking systems.
5. Transparency: All transactions are recorded on a blockchain, providing an immutable audit trail for risk management and compliance.
As AI agents become more prevalent, privacy and trust must be assured. But how can we verify that an AI agent is making the right decisions and without compromising sensitive data?
This is where ZKPs come into play.
ZKPs are cryptographic tools that enable one party (the prover) to demonstrate the validity of a statement to another party (the verifier) without revealing any information other than the fact that the statement is indeed true. ZKPs are trustless because the verifier does not need to assume the prover or any intermediaries are acting honestly to verify the proof.
ZKPs accordingly allow the properties and operations of AI agents to be verified without revealing the underlying data or algorithms. ZKPs can be used to verify that an agent’s model is genuine and authorized to make transactions.
ZKPs can also demonstrate that an AI agent is operating within regulatory boundaries without exposing sensitive operational details. For instance, an AI system managing USDC-based remittances could prove it's adhering to anti-money laundering regulations without revealing individual transaction details.
ZKPs are a mathematical way to represent computations that can be verified in a trustless manner. Since AI agents themselves are computer programs, those programs can be turned into statements that can be run through the verifiability math of ZKP.
By verifying equations about AI agents, we have strong guarantees that the agent did what its programming meant it to do. The zero knowledge aspect is that private inputs can be used to verify AI agents.
At NovaNet, we built a ZKP system well suited for verifying AI agents. NovaNet ZKPs are optimized for portability in the many environments that AI agents operate, including consumer GPUs and edge devices.
They are portable because we focus on memory efficiency with folding schemes. Folding schemes are recursive ZKP processes that allow users to pick the computational step size that matches the hardware needed for their use cases. Large machines have bigger step sizes and run very fast, while smaller machines can have smaller step sizes and run slower.
Depending on the AI program, the model can be executed locally or executed on an untrusted computationally powerful prover. Generating proofs locally enables privacy because the data stays on the device or network where the private data is located. Using third parties for proof generation enables more general verifiable compute but without privacy.
One of the most exciting applications of ZKPs is the ability to verify off-chain computations that trigger USDC payments.
Verifying off-chain compute is crucial. Most AI agent operations take place off-chain due interacting with off-chain systems. AI compute is also usually too heavy to be carried out on chain.
When it comes to off-chain compute and AI agents, there are three main ways for ZKPs to trigger USDC payments:
1. Verifying Complex Calculations
ZKPs can generate proofs for complex off-chain calculations performed by AI models. This enables an AI agent to make a USDC payment based on the verified decision.
For consumer apps, AI agents that power self-driving taxis could calculate fares and make USDC payments to fleet owners without revealing sensitive route or passenger information. Likewise, to incentivize group fitness, health trackers can transfer USDC upon completion of tasks without exposing personal wellness data.
For supply chain, an AI agent could trigger USDC payments to suppliers when stock levels reach certain thresholds, while proving the decision is based on accurate inventory data without revealing sensitive sales information.
In decentralized finance, an AI-managed yield farming bot could prove it's generating USDC returns above a certain threshold without revealing its exact strategy or the pools it's utilizing.
2. Verifying AI Outputs and Models
ZKPs can verify the outputs and models of AI inference without revealing the underlying model or data. This means that ZKPs can prove that an AI agent’s decision meets predefined requirements without revealing the exact inference process.
This can be done because AI inference itself can be run in a ZKP prover to verify that the AI ran accurately and in compliance without exposing the data or model that was used.
Verifying models and outputs means that an AI shopping assistant could prove it's finding the best deals without sharing a user's browsing history or purchase patterns. Also, AI agents can use an insurance platform to assess claims and trigger USDC payouts with proofs that each decision follows AI-driven policy terms.
3. Verifying Data While Keeping Privacy
ZKPs can ensure the integrity of data used in computations without exposing the raw data. This enables proving that the AI made a decision based on accurate data without revealing the data itself.
For instance, an AI trading bot could prove it's making decisions based on real-time market data without exposing its data sources.
Another example is verified biometric data on a phone used in USDC workflows. Memory-efficient NovaNet ZKPs can run on a phone to verify the model and biometric data. We call this personally verifiable AI because the model runs locally and is portable for verification by a wide variety of third parties.
NovaNet’s open source zkEngine is a zero knowledge virtual machine that can be used to trigger USDC payments by a verifiable AI agent. If the proof is verified, then the verify method will return a successful result. This successful result can trigger a USDC payment by an AI agent by integrating the USDC smart contract.
As discussed above, an AI agent may require ZKP verification to qualify to make USDC payments. In this simple example (in Rust), a Gradient Boosting Machine AI model can be run through the zkEngine for verification.
`fn main() -> anyhow::Result<()> {
init_logger();
// Configure the arguments needed for WASM execution
// Here we are configuring the path to the WASM file
let args = WASMArgsBuilder::default()
.file_path(PathBuf::from("wasm/gradient_boosting.wasm"))
.invoke(Some(String::from("_start")))
.trace_slice_values(TraceSliceValues::new(0, 100_000))
.build();
// Create a WASM execution context for proving.
let mut wasm_ctx = WASMCtx::new_from_file(args)?;
// Prove execution and run memory consistency checks
// Get proof for verification and corresponding public values
// Above type alias's (for the backend config) get used here
let (proof, public_values, _) =
BatchedZKEProof::<E1, BS1<E1>, S1<E1>, S2<E1>>::prove_wasm(&mut wasm_ctx)?;
// Verify proof
let result = proof.verify(public_values)?;
Ok(assert!(result))
}
}
`
When it comes to payment security for AI agent transactions, a particular benefit of ZKPs is improving escrow systems.
ZKPs provide trustless proof that funds are locked in escrow. As explained by Circle’s Corey Cooper, ZKPs enable trading “escrow services [to] run on smart contracts rather than through centralized order books.” Importantly, ZKP escrow services can be provided without revealing more information than parties need to disclose. This is particularly useful where maintaining privacy about the transaction value or identities of the parties is crucial.
In addition, ZKPs can verify that conditions for release of funds from escrow have been met without disclosing the specific conditions. Complex smart contracts involving multiple AI agents or intricate conditions can remain confidential. For example, parties in a supply chain might not want to disclose the specific quality standards or inspection results due to competitive or proprietary concerns.
The convergence of AI agents, USDC, and ZKPs is paving the way for a revolutionary autonomous and trustless global payment system. This synergy promises to unlock new possibilities in financial operations, offering unprecedented levels of verifiability and privacy in transactions across borders and time zones.
