# Circle details Anduin payment-graph reputation test for AI agents

Published 2026-11-13 · Crypto.world Team · Crypto.world

> Circle described Anduin, an experimental system that uses agents’ onchain payment histories to rank service providers. In a study of 813 buyers, its amplified-flow method cost $8,259 to manipulate across all buyers, compared with $358 for collaborative filtering.

## Payment histories as reputation

[Circle](https://crypto.world/companies/circle) has detailed [Anduin](https://arxiv.org/abs/2611.05586), an experimental system that ranks service providers for AI agents using the payment relationships visible in onchain records. Circle described the approach in a post dated Nov. 13; the underlying research paper was submitted to arXiv as a preprint on Nov. 3.

The paper models payments as a graph linking buyers to providers. A provider’s rating for a buyer is calculated from graph paths connected to services that buyer already pays; the proposed method uses payment flows rather than voluntary reviews or simple popularity counts. The researchers tested three variants, called maximum flow, summation flow and amplified flow.

## Study results and attack costs

The dataset covered 157 days, from March 23 to Aug. 26, 2026, and included 66,073,201 x402 payments across 16 blockchains, with $6.1 million in volume. The paper reports 162,882 buyers and 50,390 providers in the full dataset, with a median payment of $0.01. For its ranking test, the researchers used a 90-day payment-history window and examined 813 buyers who later adopted a previously uncontacted provider already paid by at least two buyers.

In a modeled attack to push a new provider to the top rank for all 813 buyers, amplified flow required $8,259 in spending, equal to 20% of those buyers’ combined spending in the study window. The comparable collaborative-filtering method cost $358, or 0.87% of the buyers’ combined spend. Circle characterized the first attack as roughly 23 times more expensive, while the paper’s figures show amplified flow’s top-1 precision at 22.8%, against 23.2% for collaborative filtering.

The model counts money paid to providers the attacker does not control, plus a $0.017 fee per transaction. Circle says fake wallets paying one another do not raise a provider’s score under the approach, because the relevant signal comes from payments to services already used by the buyer.

## Limits and status

The work evaluates whether a ranking identifies a buyer’s next provider; it does not establish whether that provider performed well. Its guarantees also assume attackers do not control the buyer’s existing providers, according to the paper. Circle’s post says the research describes technology that is not a completed or released software product or service.

## Sources

1. Original source · Preprint [Anduin: A sybil-resistant agentic marketplace reputation system](https://arxiv.org/abs/2611.05586) arXiv:2611.05586

Other pages read

2. [Circle](https://www.circle.com/blog/sybil-resistance-in-reputations-for-agentic-marketplaces) circle.com

## Companies in this story

- [Circle](https://crypto.world/companies/circle)

## Related stories

- [Researchers Find 9 of 10 AI Models Raised Their Own Circle Wallet Caps](https://crypto.world/news/ai-models-raised-circle-wallet-spending-caps)
- [Circle Details CCTP Routes for Non-USDC ERC-20 Tokens](https://crypto.world/news/circle-cctp-non-usdc-erc20-token-routes)
- [Hyundai Card and JPMorgan test AI treasury agents](https://crypto.world/news/hyundai-card-jpmorgan-test-ai-treasury-agents)
- [Circle Names Disparte Vice Chairman, Adds COO Role for Walia](https://crypto.world/news/circle-disparte-vice-chairman-walia-chief-operating-officer)
- [World Bank explores USDC settlement with Citi and Circle](https://crypto.world/news/world-bank-explores-usdc-settlement-citi-circle)

## Explore

[More articles](https://crypto.world/news)

---

Canonical: https://crypto.world/news/circle-anduin-payment-graph-agent-reputation-test
