Most security studies focus on the fires. This one maps the fuel.
A USENIX Security '26 paper dropped 65,340 addresses into the public ledger. 126,982.94 ETH and 17,726.7 BNB. Valued at $574.8 million using May 2025 reference prices of $4,408 per ETH and $847 per BNB. The researchers call them "risky crypto addresses." I call them infrastructure debt that someone will eventually collect.
The headline number is staggering. But the real story is the two active attack vectors that directly account for only $15.7 million—2.7% of that total. The remaining 97.3% is a time bomb sitting on mainnet, waiting for either a deployer to show up or a sweeper to find the right key.
Code is law, but audit is mercy. And this dataset has not been audited for mercy.
Context: How the Dataset was Built
The researchers did something few have the patience to do. They mined 63,004 GitHub repositories created between January 2015 and May 2025. They extracted 16.3 million deduplicated private keys. Derived addresses. Combined direct key matches with transaction-pattern rules and lightweight symbolic execution on Ethereum and BNB Smart Chain.
That's forensic work. The kind I did on the 2x Funding contracts in 2017, when I found an integer overflow in leverage calculation logic that would have drained user funds during volatility. The difference is that my audit focused on a single contract. They scanned an entire ecosystem.
They reported 99.11% precision for their overall address-misuse detection. That means out of 100 flagged addresses, roughly 99 are genuinely risky. Two researchers independently judged random samples. The precision figure measures detection accuracy, not the dollar value directly attributable to the two active vectors.
But here's the rub: the $574.8 million figure uses May 2025 prices. The losses occurred across years. Some ETH was worth $200, some $4,800. The economic impact is not a static number. It's a moving target that depends on when you mark to market.
Core: The Two Active Vectors and Why They Matter
The study separates misuse into two categories: contract-account misuse and externally owned account (EOA) misuse.
Vector 1: Testnet-to-Mainnet Contract Addressing
Contract-account misuse occurs when someone sends a function call, sometimes with ETH or BNB attached, to an address that has no contract code on the selected network. The transaction succeeds as a simple transfer. Funds sit at that address. If later-deployed code can move them, they become accessible.
An attacker deploys a contract on a testnet. Waits for users to mistakenly send funds to the corresponding no-code address on mainnet. Then deploys malicious withdrawal code at the same deterministic address. The paper identified 469 malicious contracts tied to 3,446.37 ETH and 431.79 BNB in losses.
This is not a novel attack. I saw similar patterns in 2019 during the DeFi composability audits I led for Compound. The difference is scale. The researchers found 469 contracts. That's a curated list of successful exploits. The total number of testnet-to-mainnet collisions is likely much higher.
Vector 2: EIP-7702 Delegation
EOA misuse starts with a public or otherwise exposed private key. Anyone who has the key controls the account. Automated sweepers race to remove incoming funds.
EIP-7702 makes that drain more direct. An attacker uses the exposed key to delegate the account to malicious code that forwards a deposit to the attacker in the same transaction. The analysis identified more than 17,200 delegated addresses. Losses: 25.86 ETH plus 33.45 BNB.
Seventeen thousand two hundred delegated addresses. That's a surface area large enough to map a city. Yet the cryptocurrency community reacts to this with the same shrug it gives to Tether's unaudited reserves.
Composability is leverage until it is liability. EIP-7702 was designed to improve user experience. It also created a new attack surface for exposed keys.
Together, the two vectors account for 3,472.23 ETH and 465.24 BNB. The rest of the aggregate covers the broader set of detected contract-account and exposed-key misuse.
Contrarian: The Industry's Blind Spot
The conventional wisdom says the solution is better user education. "Check the address. Check the chain. Don't paste private keys into GitHub."
That's like telling people to stop breathing. The structural problem is that blockchain infrastructure codifies these mistakes permanently.
Deterministic contract addressing is a feature. It allows predictable deployments. But it also creates a permanent link between testnet experiments and mainnet funds. The paper's 469 malicious contracts are just the ones we know about. How many other testnet addresses on Ethereum mainnet hold funds that no one has claimed?
Exposed private keys on GitHub are not a user error. They are a system design failure. The blockchain does not distinguish between a key that was intentionally shared and one that was accidentally committed. The protocol treats both as valid.
"Trust no one, verify everything, build twice." That's my mantra. But the industry builds once, ships fast, and hopes the auditors catch the rest.
Here's the uncomfortable truth: the 99.11% precision reported by the researchers is impressive, but it measures detection, not remediation. The paper does not provide a complete remediation rate for all 65,340 instances. They began disclosing to wallet developers and exchanges. But disclosure is not the same as fixing.
When I consulted for the consortium evaluating BlackRock's ETF infrastructure, we spent weeks on Arbitrum's fraud proof mechanisms. Gas cost savings of 90% compared to L1. Finality from 7 days to 24 hours. But we also spent equal time on key management. The infrastructure is only as strong as the keys that control it.
The study's $574.8 million figure is a static snapshot. The real economic damage is dynamic. Each of those 65,340 addresses is a potential future exploit. The dormant ones are the most dangerous. They have no current activity, so they escape detection. But when a new vulnerability emerges—a new EIP, a new contract standard—those dormant addresses become active attack surfaces.
Infinite yield curves break under finite scrutiny. The same applies to security. The industry has finite attention. The attackers have infinite time.
Takeaway: The Responsibility Lies with Infrastructure
Users can reduce immediate risk by checking both the address and chain against official sources. Developers should keep test accounts and hardcoded keys out of production. Wallet providers can warn before transactions reach no-code or exposed-key destinations.
These are tactical fixes. Strategic fixes require changing how the protocol handles address creation and key management. Maybe we need address-based warnings integrated into wallets. Maybe we need a standard for testnet-to-mainnet separation that is enforced at the consensus level.
Blind faith is the only true vulnerability. The 65,340 addresses are evidence of past mistakes. But they are also a forecast of future ones. The question is not whether the remaining $559 million will be lost. The question is how many more addresses will be added to the list before the next USENIX paper.
Logic dictates value, perception dictates volume. The value of this study is not in the $574.8 million headline. It's in the map of the fuel. The fires will come.
