Crypto exchange Bybit said AI-assisted auditing found high-severity vulnerabilities at up to five times the rate of manual review, cutting the time between assessing a system and testing it from about two weeks to two hours.
Between Jan. 1 and June 15, Bybit said the AI system also stopped more than 30,000 suspicious withdrawal requests, protecting close to 20,000 users from over $700 million in potential losses. The first review of a flagged withdrawal averaged 4.7 minutes, the exchange said in a release shared with CoinDesk.
There’s a reason the exchange is publishing figures like these. In February 2025, it lost roughly $1.46 billion in the largest theft in crypto's history, attributed to North Korea's Lazarus Group. It is now pursuing legal action against the state and the group.
"Potential losses" are withdrawals Bybit says it stopped rather than thefts that were definitely underway. AI also flagged about $212 million in funds it linked to fraud and blacklisted more than 10,000 addresses, the company said, though the figures cannot be verified independently.
The firm’s automated red-team system scanned 1,489 public-facing assets in the period and turned up more than 100 high-severity flaws. The time between finding an asset and testing it fell below 24 hours. More than 100,000 security alerts were processed with AI assistance over the period.
Such AI-assisted protections come in a month that has seen smaller firms and independent developers turn the same tools on their own products, hunting for flaws before attackers find them.
BTCPay Server, which earlier this month suffered an attack that drained merchant Lightning nodes, said AI is changing the balance between attackers and defenders. Artificial intelligence models make it faster and cheaper to search large codebases for weaknesses, and attackers, especially state-sponsored teams, may be able to call on greater financial resources.
Last week, dozens of crypto-aligned firms, including Coinbase COIN$146.18·마감 시 and Block XYZ$79.63·마감 시, signed an open letter asking AI labs for early access to their strongest models, arguing defenders are working with weaker tools than the people attacking them.
Separately, a volunteer group calling itself the Bitcoin Red Team has spent the month pointing AI models at bitcoin codebases and filing thousands of findings across hundreds of projects, including a report that helped BTCPay patch it weakness.