Hook
In my five years tracking MEV, I've learned that the most dangerous anomalies often hide in plain sight. Last week, while cross-referencing on-chain data, I stumbled upon a number that made me pause: 6.85 million sandwich attacks on Solana over an 11-month span, extracting $29.7 million. That's an average of 20,600 attacks per day, or roughly one every 4.2 seconds. But the real shock wasn't the frequency—it was the concentration. 96% of these attacks targeted Pump.fun's bonding curve and PumpSwap. This wasn't a random exploit; it was a systematic tax on retail traders.
Context
Sandwich attacks are a well-known MEV tactic. A bot spots a pending trade, front-runs it with a buy, then back-runs it with a sell, profiting from the price impact. On Ethereum, this happens in the public mempool, where bots compete to bribe miners. Solana was supposed to be different. Its high throughput, low fees, and lack of a public mempool theoretically made sandwiching harder. Yet here we are, with 2,545 bot wallets extracting value from over 2.15 million unique wallets. The study, a preprint from an anonymous author, claims to cover October 30, 2025 to October 7, 2026—a window that ends after the preprint's submission date of September 23, 2026. That's a red flag I'll return to. But the underlying technical reality—that bonding curves are a magnet for predatory MEV—is undeniable. Follow the gas, not the hype: the data shows Solana's MEV problem is not a bug, but a structural feature of its AMM design.
Core
Let's dig into the mechanics. The study identifies two distinct bot strategies based on slippage tolerance. Transactions with 20–50% slippage were attacked 14 times more frequently than those with <3% slippage. But here's the twist: trades with <0.5% slippage were attacked more than those with 1–3% slippage, and trades with >50% slippage were attacked less than the 20–50% band. This nonlinear pattern suggests a layered bot ecosystem: one class of bots hunts large, wide-slippage trades for big margins, while another class executes high-frequency, micro-profit attacks on tiny slippage. The 1–3% range is likely an unattractive middle ground—too risky for micro-bots, too unprofitable for large ones.
This is classic market segmentation. But the real story is the economic sieve. The study reports that bots retained 68% of gross revenue in a sample of 63 attacks, with 28.6% of those attacks actually losing money after costs. That means roughly 32% of gross profits are eaten by gas, priority fees, failed retries, and Jito tips. The $29.7 million figure is a gross extraction, not net profit—anyone citing it as 'bot windfall' is missing the cost structure. More importantly, 39% of extracted value came from other bots. Yes, bots are sandwiching bots. This is toxic MEV, where high-priority bots front-run lower-priority ones. The narrative that retail traders are the sole victims is incomplete; a significant chunk of value is transferring between machines.
Now, the Pump.fun connection. Bonding curves are predictable by design. The price follows a deterministic formula, so any trade large enough to move the curve is a sitting duck. This is why 96% of attacks cluster there. Pump.fun's fee accounting is also inconsistent—the study notes a 3% discrepancy in sell fee records. That's a hidden risk for $PUMP valuation models, as it suggests systematic underreporting of revenue. And Jupiter? It only carried a small fraction of victim trades, implying that its routing or default slippage settings offer some protection. Check the supply, trust the chain—but also check the fee records, because they don't always add up.
Contrarian
The most counterintuitive finding is that Solana's speed and low fees may actually facilitate sandwiching. Without a public mempool, bots rely on Jito bundles or leader ordering to position their transactions. The study admits that on-chain records cannot prove bots 'saw' the order in advance—they could be backrunning based on intra-block state. This legal gray area makes enforcement nearly impossible.
But here's the bigger blind spot: the data itself is suspect. The preprint's date anomaly—submission before the data window ends—is a hard logical flaw. Either the dates are wrong, or the study is fabricated. Until peer review, we must treat the $29.7M as a lower bound with unknown error bars. Whales move in silence. Listen closely—but also question the timestamp.
Takeaway
The real signal here isn't the dollar amount; it's the structural vulnerability of bonding curves to MEV. As long as Meme season persists, sandwich bots will feast. The next time you see a token launch on Pump.fun, ask yourself: who's on the other side of your trade? Probably a bot. And if you're using a DEX aggregator, check your slippage settings—they might be your only shield. Liquidity leaves first. Panic follows. But data? Data stays.
