6 min read

Kelp DAO rsETH Exploit: What Actually Happened

Source: http://x.com/i/article/2045802734306123776

~$292M drained. 116,500

**$rsETH** stolen. The largest DeFi exploit of 2026 so far.

There’s a lot of rumor circulating on CT – and the reality is NOT what they describe it. Let me synthesize it clearly, separating what is proven onchain from what remains uncertain.

PS: For a deeper dive & evidence, make sure to check this investigation: https://gist.github.com/banteg/705d0284513b74ad20f61d90f5b5de62!

[Image](https://pbs.twimg.com/media/HGQm6H6XkAAK6dbformat=jpg&name=large) # TL;DR: What ACTUALLY happened

A structurally normal-looking LayerZero packet (nonce 308) was verified, committed, and executed on Ethereum — releasing 116,500 rsETH from real bridge inventory —  without any corresponding burn, debit, or nonce advance on Unichain.

This was not a mint exploit. It was not a reentrancy. It was a  phantom cross-chain message that the Ethereum side accepted as legitimate.

Now, let’s have a closer look

1. The Route was a single point of failure

The Unichain → Ethereum Kelp path was configured as a  1-of-1 DVN (Decentralized Verifier Network):

Image

This is the critical architectural weakness.

In LayerZero's security model, you can configure X-of-Y-of-N verifiers. Kelp used  1-of-1 — meaning a single verifier compromise is sufficient to forge any message on this route.

Importantly, this was  not a last-minute config change. The 1-of-1 setup was in place at least 90 days before the exploit. It was a longstanding design choice.

So should we now blame Kelp for negligence? The situation is much more complex.

First of all, the sole DVN was LayerZero's own infrastructure.

This is the part that fundamentally changes the blame calculus. The required DVN on the Ethereum side (0x589ded...) is  labeled "LayerZero: DVN" on Etherscan . This wasn't some random third-party verifier Kelp chose to trust blindly — it was LayerZero's own DVN product.

When you configure 1-of-1 with LayerZero Labs as the sole verifier, you are implicitly trusting  LayerZero's own operational security. That's a very different risk profile than trusting an unknown external party. The reasonable assumption is: "If LayerZero's own DVN fails, that's LayerZero's problem to disclose."

Secondly, 1-of-1 with a trusted provider is an industry-common pattern.

Look at the broader context from the investigation:  11 out of 12 Kelp inbound routes to the Ethereum adapter were configured as 1-of-1.

This wasn't a Unichain-specific oversight — it was Kelp's standard deployment pattern across all chains. And notably,  only Unichain shows a nonce mismatch. Every other route is operating normally with the same config.

That asymmetry matters.

If the 1-of-1 config itself were the root cause, you'd expect to see anomalies across all routes. The fact that only Unichain was exploited suggests the failure was  route-specific or DVN-operator-specific, not purely a config-level weakness.

2. The Phantom Packet (Nonce 308)

The attack unfolded in three tight steps over ~2 minutes:

The packet was  structurally indistinguishable from a legitimate one. Same origin tuple, same receiver, same 28-byte payload shape as the clean control packet (nonce 307). No malformed encoding. The only anomaly was invisible to the Ethereum endpoint: there was no real source-side event.

3. The smoking gun: the nonce mismatch

This is the clearest proof of the exploit's nature:

Ethereum destination endpoint: lazyInboundNonce = 308 ✅ (accepted nonce 308) Unichain source endpoint: outboundNonce = 307 ❌ (never sent nonce 308)

For every other Kelp route (Arbitrum, Base, zkSync, Mantle, Scroll, etc.), source and destination nonces match.

Unichain is the only route with a true nonce mismatch.

The Unichain source token's totalSupply() also remained completely flat — no 116,500 rsETH burn ever happened.

4. The executor was a throwaway mixer wallet

LayerZero's model separates verification from execution — after a packet is verified, execution is  permissionless.

The attacker exploited this:

  • The  configured executor (0x4208D6...) was dormant — nonce 0, no code, no activity
  • The  actual executor (0x4966260...) was a fresh wallet with exactly 5 txs, all on this Kelp route, funded hours earlier via  Tornado Cash (0x47ce0c...)

This is a deliberate design: use LayerZero's permissionless post-verification execution to deliver the packet yourself, without needing to compromise the executor role.

5. The DVN signer committee was stable — but admin surface expanded

The destination DVN (0x589ded...) is  not a proxy — no upgradeable implementation slot. Its signer committee (3 signers, quorum 2) was unchanged around the exploit window. No UpdateSigner or UpdateQuorum events.

However, what did change:

  • Two addresses (0xD3A227... and 0xD95D77...) were granted ADMIN_ROLE in  February 2025
  • A broader admin expansion occurred on  April 8, 2026 — 10 new grantRole(ADMIN_ROLE, ...) txs, 10 days before the exploit

Critically, these same two addresses were the ones that submitted PayloadVerified for nonces 308 and 309. They are  long-standing authorized DVN admins, not freshly injected attacker keys.

They also appear to be verifying many other packets on the same receive library — consistent with routine DVN operator behavior.

What the attacker did with the funds

The 116,500 rsETH was dispersed in ~5 minutes into 7 pre-staged branch wallets, all Tornado-funded.

Here's the breakdown:

Image

Ethereum collector (0x5d3919...): ~75,700 ETH received,  nonce 0 — hasn't moved yet. Arbitrum collector (0x5d3987...): ~30,765 ETH equivalent received, nonce 6, with some unusual zero-value txs back to branch wallets.

The open Aave positions (both Ethereum and Arbitrum) remain live:

Image

These positions are dangerously close to liquidation thresholds — and they represent the attacker's  ongoing exposure on Aave.

What Kelp did right (and fast)

  • 18:23:11 UTC — Kelp's Safe multisig froze recipient 0x8b1b6c... via sweep(address[],address) on the rsETH token, setting transfersBlockedUntil = 2026-04-19 18:23:11 UTC
  • This blocked nonce 309 (a second phantom packet for 40,000 rsETH) from executing — even though it had already been verified and committed
  • Nonce 309 reverts with TransfersBlocked(0x8b1b6c..., ...) — the adapter had sufficient inventory (40,357 rsETH), so the block was the only thing stopping a second drain
  • Paused rsETH contracts on mainnet and multiple L2s

The freeze saved ~$100M.

The 46-minute response window was tight but effective.

Important note in this context is that all rsETH on mainnet remains fully backed. The  underlying LST collateral (stETH, ETHx, etc.) held in Kelp's core protocol contracts on Ethereum mainnet was  never touched by the exploit.

What remains unknown

The exact root cause inside the verification path is  not yet isolated:

Hypothesis 1: LayerZero DVN was compromised

  • Evidence for:

DVN is LZ-labeled; admin expansion 10 days prior; submitters are long-standing admins

  • Evidence against:

Signer committee unchanged; same admins verifying many other packets normally

Hypothesis 2: DVN had a verification bug

  • Evidence for:

Packet was structurally normal; no source-side event needed to trigger it

  • Evidence against:

No other routes show nonce anomalies

Hypothesis 3: Bad upstream input on Unichain side

  • Evidence for:

Source nonce never advanced; no burn event

  • Evidence against:

Route config was stable for 90 days

The fault domain is  narrower than "some weak third-party DVN" — the sole required DVN is Etherscan-labeled as LayerZero-associated. But it's  wider than a confirmed LZ infrastructure breach — the same DVN is operating normally on hundreds of other routes.

What CT is getting wrong

Most narratives circulating are oversimplified. Here's the correction matrix:

Image

Broader implications For LayerZero users:

The 1-of-1 DVN configuration is explicitly supported and documented by LayerZero as a valid minimal setup. But this incident demonstrates that it creates a single point of failure — whether through compromise, bug, or bad input. The recommended posture is multi-DVN with independent providers.

For Aave: The protocol is now sitting on an estimated ~$177M in bad debt from rsETH collateral positions that cannot be cleanly liquidated without crashing rsETH further. The health factors on the attacker's open positions (1.027–1.032) are precariously close to liquidation.

For the broader ecosystem:

This is 2026's largest DeFi exploit. The attacker's ETH proceeds (~75,700 ETH on Ethereum alone) are sitting idle in a nonce-0 collector wallet. The next move — whether that's mixing, bridging, or OTC — will be closely watched.