The research Alba stands on

Each source below is listed with what it found and with what Alba does about it.

Agents leak what their owners know

AgentCIScope: Grounded Contextual Integrity Evaluation for LLM Agents

Guo, Chen, Lin, Liu, L. arXiv 2610.19158, 13 Oct 2026

Found

Across 12 agents and 936 multi-step tasks, the principal’s private context leaked in 32.5% to 94.4% of single runs.

In Alba

A grant credential opens one grant. The balance, the other grants and every ledger stay behind the owner’s key, which the agent never holds, so there is nothing more for it to give away.

Got a Secret? LLM Agents Can't Keep It

Priyanshu, Vijay, Pahw. arXiv 2605.27766, 26 May 2026

Found

Leakage stayed above 37.8% even when agents were instructed to keep the information private.

In Alba

Alba does not ask the agent to keep a limit. Each term of a grant is a constraint inside the proof, and a payment that breaks one has no valid proof.

What an x402 payment shows

How Agentic Is Agentic Commerce?

Ling, Zhou, Wu, Wan. arXiv 2607.12575, 14 Jul 2026

Found

Over 280 days on Base, 136,708,672 settlements worth $44.12M were clustered by address. Payers were linked to each other and to the services they paid, from public data alone.

In Alba

The payer on record for every Alba payment is the pool. There is no payer address to cluster.

ERC-8004 measurement study

arXiv 2606.26028, 24 Jun 2026

Found

The same payment graph ties settlements to registered agent identities.

In Alba

A grant’s payments share no public identifier. Each one spends a fresh tag and leaves a fresh commitment.

Anduin

arXiv 2611.05586, 3 Nov 2026

Found

Agents are scored by whom they pay, using the public payment graph.

In Alba

Whom an agent pays is in its grant’s ledger, and the ledger opens with the owner’s key.

Found

A payment’s metadata (the resource URL, its description, reason strings) reaches the seller and a central facilitator before settlement, and neither is usually bound by a data agreement.

In Alba

The resource the agent names is kept as a memo encrypted to the owner. The facilitator verifies a proof and relays a transaction, and holds neither funds nor keys.

When HTTP 402 Meets the Blockchain: Risks on Emerging x402 Payments

Wang, Yang, Chen, Ji, Paye. arXiv 2607.19545, 21 Jul 2026 (USENIX Security 2026)

Found

15 facilitators serving over 60,000 sellers and 360,000 buyers were tested, and every one broke at least one of eight security rules.

In Alba

The program verifies every proof itself. A facilitator can refuse to relay a payment; it cannot move a grant outside its terms.

x402 on Solana

Found

x402 had settled about 130 million transactions by May 2026, 35% of them on Solana. USDC carried 98.8% of the dollar volume on EVM chains.

In Alba

Alba is built where agents already pay, and its first pool holds USDC because sellers price in it. A pool opens for any SPL mint.

x402 specification v2.0

Coinbas. 9 Dec 2025

Found

Version 2 of the x402 specification made schemes an extension point: a new way to pay stands beside exact, and facilitators advertise what they support.

In Alba

Alba is an x402 scheme named alba. A seller takes it with one line of middleware and is paid in the asset they priced in.

groth16-solana, README benchmarks

Light Protoco. github.com/Lightprotocol/groth16-solana

Found

One Groth16 verification on Solana costs about 78,000 to 109,000 compute units through the alt_bn128 syscalls.

In Alba

Each payment carries one proof and is verified inside one transaction, so a seller waits for settled money before serving a single request.

Solana program 49hapF2ZzcSCmVDqT56QTFb3UDW28HaSDqaLSjqW5YvV@usealbalive on XEvery time on this site is UTC.