A student audit of Omo, a Solana (SOL) AI trading agent that showed a portfolio above $200,000, found none of its 174 order commitments matched a trade its wallet signed.
Key Points:
- A student researcher checked one sealed Omo decision and found the paired transaction was a token transfer signed by another wallet.
- Omo's own public data tied none of its 174 order commitments to a trade the agent signed.
- The paper urges hardware-held signing keys, though its author says hardware would not have fixed the agent's other failures.
Omo Audit Findings
Charlie Sneed, a member of the University of Oregon's blockchain club, published the paper on Sept. 28 through a research competition run by hardware wallet maker Ledger. Ledger says it does not vouch for his conclusions. Sneed began studying Omo while its public dashboard showed a fast-growing portfolio, days before the agent stopped trading and its site went offline.
Omo launched on Aug. 9 and traded memecoins through Aug. 31. Before each order it posted a cryptographic fingerprint of its reasoning to the Solana blockchain, then published the text 20 minutes later so anyone could compare the two.
Its documentation lists four checks an outsider can run with a hashing tool and a public Solana server, without access to Omo's systems. The fourth asks whether the agent's published wallet signed the trade, and the project calls a fill signed by any other key disqualifying. That check failed.
Sneed verified one commitment by hand, the oldest with a trade attached. The transaction paired with that decision, made Aug. 21, was a token transfer signed by another wallet, sent to several addresses including Omo's, with no purchase in it.
Also Read: Tesla Robotaxis Get An 11 P.M. Cutoff, And Musk Has A Pet Problem To Solve
Sneed On Keys
Sneed wrote that the cryptography held up, that every problem he found was documented somewhere in Omo's code and that he saw no evidence of deliberate misrepresentation. The zero-of-174 tally comes from Omo's own public data pages. The weak point, he argued, was custody, because Omo used a key copied out of a phone app and stored on a server, and its trade-matching code never checked signatures.
Sneed proposed three fixes, led by a signing key that the agent controls and that sits in hardware, the kind of product Ledger sells. He limited that claim himself, writing that hardware would not have caught Omo's 7.1% win rate or the web search that had been failing for days while the agent kept trading.
Agent Trading Risks
Larger platforms opened to trading agents this year. Binance let AI agents trade on Aug. 20, and Jeff Li, its vice president of product, said the company "really cannot see the reasoning" behind their orders. Robinhood reported more than 150,000 agentic accounts on Sept. 29 while stating that it does not supervise or audit agents, and FINRA's 2026 oversight report flagged autonomy and auditability among the risks.
Read Next: Robinhood, Binance And Coinbase Now Take Orders From AI Agents: What To Know
