5 On-Chain Patterns That Reveal Sybil Agents
The AI agent economy has a trust problem. As autonomous agents begin transacting via protocols like x402, every participant needs to answer a fundamental question: is this agent real, or is it a manufactured identity designed to game the system?
Sybil attacks — where one operator creates many fake identities to accumulate outsized influence or rewards — are the oldest trick in decentralized systems. But agents make sybils cheaper to create and harder to detect than ever before.
At DJD Agent Score, we analyze on-chain transaction patterns to assign reputation scores to AI agent wallets. Here are five behavioral signatures that reliably expose sybil agents — even when their operators try to disguise them.
Tight Cluster Rings
Legitimate agents interact with diverse counterparties. Sybil agents exist in a manufactured ecosystem — their operator controls all the wallets, so the “agents” inevitably transact within a tight, interconnected group.
We build a relationship graph and check whether a wallet's top 5 partners share significant mutual connections. When more than 50% are interconnected, the tight_cluster indicator fires.
Real-world analog: In traditional finance, this is how investigators identify shell company networks — entities that only transact with each other are likely under common control.
Symmetric Round-Trips
Real economic activity is asymmetric — an agent that provides a service collects payments; one that consumes a service pays fees. When an operator simply moves funds between controlled wallets, the amounts going A→B and B→A tend to be suspiciously similar.
When the smaller direction is within 10% of the larger for more than half of a wallet's partnerships, symmetric_transactions fires. We also detect explicit wash trading — when more than 40% of 7-day volume consists of round-trips within 24 hours.
Coordinated Creation Windows
Organic relationships develop over time. Sybil wallets are deployed in batches — created on the same day, funded from the same source, and immediately start manufacturing activity between them.
Timing is the hardest thing to fake retroactively. Once a wallet's creation timestamp is on-chain, it's permanent. This pattern becomes especially powerful when combined with Pattern 4.
Puppet Funding Chains
Legitimate agents are funded by exchanges, bridges, or treasuries — neutral infrastructure. Sybil agents are funded by the operator's main wallet, which is also the entity they'll “transact” with to build fake reputation.
This is one of our highest-confidence signals. It simultaneously caps both Identity and Reliability dimension scores. Real agents have independence — their funding and revenue come from different sources. Puppet agents have dependence.
Bot-Like Temporal Signatures
This pattern analyzes when a wallet transacts, not who with. Human-directed agents show natural variability — business hours, weekend gaps, irregular spacing. Sybil scripts run on fixed intervals with unnaturally low variance.
We measure three things:
- Inter-arrival CV: Below 0.1 = machine-like regularity
- Hourly entropy: Low entropy = activity concentrated in a few hours
- Maximum gap: Genuine agents have downtime. Sybil scripts don't.
How These Patterns Compound
No single pattern is conclusive on its own. The power lies in pattern stacking. We apply each detected pattern as a multiplicative penalty — the integrity multiplier.
A wallet flagged for tight clustering (0.55x), symmetric transactions (0.60x), and wash trading (0.50x):
0.55 × 0.60 × 0.50 = 0.165x ← 83.5% score reductionThe multiplier floors at 0.10x — we never completely zero a score, because even our highest-confidence signals carry some false positive risk. This creates a sharp separation between legitimate agents (near 1.0x) and sybils (below 0.30x).
Why This Matters
As AI agents begin operating autonomously in the x402 ecosystem, the ability to distinguish real agents from manufactured ones becomes critical infrastructure:
- Service providers need to know if a client is trustworthy before extending credit
- Protocols need to prevent sybils from farming governance influence or rewards
- Marketplaces need to surface quality agents and suppress fake ones
With DJD Agent Score now publishing to the on-chain ERC-8004 Reputation Registry, these signals are available as public infrastructure. Any protocol on Base can call getSummary() to check an agent's reputation — no API key required.
Check your agent's score via x402 micropayments, or verify on-chain at the ERC-8004 Reputation Registry:
0x8004BAa17C55a88189AE136b182e5fdA19dE9b63