Sandwich bots extracted about $29.7 million before costs from Solana trades over an 11-month period, according to blockchain data provider Bitquery. It found 96% of detected attacks on memecoin platform Pump.fun’s launch curve and its PumpSwap exchange.
The investigation, covering successful swaps, counted 6,850,144 attacks by 2,545 bot wallets affecting 2,154,994 distinct wallets between October 30, 2025, and October 7, 2026. Those addresses do not represent a count of individual people: other bots’ trades accounted for 39% of the estimated gross take.
A sandwich places a bot’s purchase and sale around another trader’s swap. The first purchase raises the pool price; the intervening buyer receives fewer tokens, and the bot sells into the higher price. The cost reaches the trader through execution, even when the swap completes successfully.
Wide slippage marked greater exposure
Among trades sent by swap router Jupiter through Pump.fun’s launch curve or PumpSwap, trades allowing 20%–50% slippage were sandwiched 14 times as often as those allowing under 3%, Bitquery found.
Slippage tolerance sets how far execution can depart from a quote. Jupiter’s documentation describes a minimum acceptable output after applying that tolerance. A price move beyond the limit can cause a swap to fail; a wider limit permits a worse fill. That creates a tradeoff between completing a volatile trade and limiting the price concession accepted along the way.
The observed attack rates did not rise uniformly with every wider setting. Trades below 0.5% tolerance were hit more often than those at 1%–3%, while the 50%-or-more band was hit less often than the 20%–50% band.
The slippage comparison is observational and covers Jupiter routes through Pump.fun’s launch curve or PumpSwap. Jupiter carried only a small share of all victim trades in the investigation.
The $29.7 million estimate values SOL at the minute of each attack and does not measure net profit. In Bitquery’s separate 63-attack fee sample, bots retained 68% of aggregate gross proceeds, and 18 attacks lost money after costs. That sample does not establish net earnings for the full period. The report also identifies inconsistent Pump.fun sale-fee recording that affects its gross estimate by about 3%.
Bitquery used repeated trading patterns to identify bots, excluded round trips beginning in a token’s first 10 slots, and reports gaps in exchange and block coverage. Chain records cannot prove a bot saw an order beforehand, and some qualifying wallets may be fast traders.
Slippage limits address the price a trader will accept. A separate research preprint submitted September 23, 2026 examined protected order flow, meaning transactions intended to be shielded from front-running, across six blockchains. Its authors found sandwich patterns extending across trades and blocks and concluded these protections can offer weaker guarantees than users expect.