BlackHart
Go live
Hacks feed/jaredfromsubway MEV bot

jaredfromsubway MEV bot

June 20, 2026·Ethereum·Logic Error
$7.5M
total loss
StatusConfirmed

Why it happened

Attack class: Logic error

Bait tokens collected approvals the bot never revoked, then one transaction used them to transferFrom the bot's WETH, USDC and USDT.

What would have stopped it

  • Reset every token approval to zero at the end of each trade.
  • Trade new tokens only through an allow-list or a wallet holding a small float.

Not scored before the exploit

BlackHart had not scored jaredfromsubway MEV bot before this exploit.

An attacker baited jaredfromsubway, Ethereum's busiest sandwich-trading bot, with dozens of fake tokens and pools built to look like profitable trades. To trade them the bot gave the fake contracts permission to spend its funds, and the attacker made sure those permissions were never used up. On June 20 one transaction used the leftover permissions to pull about $7.48 million in WETH, USDC and USDT out of the bot. No ordinary users or public protocols were affected; the proceeds were turned into ether, with 3,000 ETH sent into Tornado Cash.

Ordinary userssafe
Public DEXs and protocolssafe
Bot treasurydrained(1,474.6 WETH, 2.87M USDC, 2.04M USDT)
Exploit anatomy

Bait contracts run by a coordinator collected standing approvals from the MEV bot, then used them in one sweep to move about $7.48 million to the attacker wallet.

The attack in 4 steps. Step 1, Root control: MEV bot contract (source). Step 2, Sweep: Bait coordinator (hub), about $7.48M reached this step. Step 3, Consolidate: Attacker main wallet (attacker). Step 4, Mixer: Tornado Cash (exit).

  1. Step 1: Root control

  2. Step 2: Sweep

  3. Step 3: Consolidate

    • Attacker
      Attacker main wallet
      stablecoins swapped to ETH
      split across five wallets
      0x3e37f4a1...ea65d0
  4. Step 4: Mixer

    • Exit
      Tornado Cash
      3,000 ETH in
      about $2.44M parked in DAI

Why each step worked

    • Approvals to unknown contractsbot granted spend rights to fresh bait tokens
    • No allowance resetmissing controlunused approvals were never revoked
    • No cap on exposure to new tokensmissing controlmain treasury reachable from any route
    • No post-trade allowance checkmissing controlunspent approvals went unnoticed
    • Automated trading at machine speedno human review of new routes
Full forensic detail

Step-by-step reconstruction, root cause, counterfactuals, remediation, and disclosure timeline.

Exploit anatomy

1.
The attacker deployed helper contracts and, per CertiK, 131 bait tokens posing as wrapped or restaked WETH and stablecoins, plus fake liquidity pools priced to look like profitable trades. The bait-token deployment CertiK cites landed at 16:03 UTC on June 19, 2026; Blockaid says the setup was built over weeks.
2.
The bot took the bait at 21:49 UTC on June 19. To trade the fake routes it approved the bait contracts to spend its WETH, USDC and USDT. Early trades used those approvals in full and paid the bot a small profit, so it kept trading the routes.
3.
On June 20 the attacker switched the bait contracts into a mode where trades no longer spent the bot's approvals, and paid builders to land ahead of the bot. Across hundreds of the bot's own trades, approvals accumulated without being used or revoked. The bot never checked that an approval had been consumed.
4.
At 18:49:11 UTC on June 20 (block 25,360,696), one transaction activated 66 bait contracts, 60 of them holding approvals from the bot, and pulled 1,474.58 WETH, 2,870,573 USDC and 2,035,760 USDT, about $7.48 million, from the bot into the attacker's main wallet.
5.
The attacker swapped the stablecoins for ether and spread the proceeds across five wallets. Per CertiK, 3,000 ETH went into Tornado Cash and about 1,422 ETH was swapped for roughly $2.44 million of DAI, which sat in one wallet.

Root cause

The bot trusted contracts it had never seen before. jaredfromsubway's searcher automatically traded any route its models judged profitable, including routes through freshly deployed tokens and pools, and to do that it granted those unknown contracts permission to spend its WETH, USDC and USDT. It assumed a granted approval would be used up inside the same trade, never checked that it was, and never revoked what was left. The attacker built tokens that behaved honestly while the bot was learning the route, then flipped them into a mode where trades left the approvals untouched, so allowances quietly piled up over hundreds of the bot's own transactions. One coordinated transaction then used the standing approvals to pull about $7.48 million. No private key was stolen and no bug in a public protocol was used; Blockaid and CertiK both describe it as an attack on the bot's decision logic. The bot's speed and automation were the weakness: it gave spending rights to attacker code at machine speed with no allow-list, no per-trade allowance reset and no cap on its exposure to any single new token.

Prevention analysis

Reset every token approval to zero at the end of each trade.

Approvals that never outlive a trade cannot accumulate. The final sweep would have found nothing to spend.

Trade new tokens only through an allow-list or a wallet holding a small float.

Routing unknown tokens through an isolated contract with a small balance would have capped the loss at that float instead of the bot's main treasury.

Check after each trade that the approval was actually used.

The trap relied on trades that left approvals unspent. A post-trade allowance check would have flagged the first such trade and stopped the bot.

Similar incidents

MEV bot 0xbaDc0dE (2022)

An attacker made an MEV bot act on attacker-supplied input, through an unprotected callback, and drained about 1,100 ETH. Same class: a searcher that executes or approves what outside contracts hand it.

Sandwich bots and a rogue validator (2023)

Sandwich bots lost about $25 million when an attacker exploited their assumptions about how bundles would execute. Same theme: automated MEV strategies baited by an adversary who studied them.

Remediation

1.Pending: the bot's operator has not published a post-mortem or said what controls were changed.pending
2.Searchers should reset approvals after each trade, isolate trades in new tokens from the main balance, and cap exposure to any newly seen contract.
3.Security firms tracked the attacker's wallets; per CertiK about $2.44 million in DAI remained in a single wallet after the drain.

Timeline

2026-06-19Attacker deploys bait token contracts.
2026-06-19Bot makes its first trade through a bait route; the approval is used normally.
2026-06-20Attacker configures the coordinator that controls the bait tokens.
2026-06-20Attacker starts paying to land ahead of the bot while approvals build up.
2026-06-20One transaction activates 66 bait contracts and pulls about $7.48 million from the bot.
2026-06-20Blockaid publishes an alert on the drain.
2026-06-25CertiK publishes its incident analysis and fund flow.
Blockaid. CertiK
Continuous adversarial monitoring

Get your protocol scored across 12 dimensions, or request ongoing coverage.