N-01 · post-mortem

The Kelp Incident.

restaking · LRT · contagion · 20 Apr 2026

$13 billion in 48 hours. A close-read of the April 2026 contagion.

The DeFi system did not break because a single protocol failed. It broke because the architecture made one protocol's failure equivalent to everyone's. This piece walks through the cascade from the initial exploit at KelpDAO's restaking layer to the $13.2 billion drawdown CoinDesk recorded by the morning of 22 April 2026 — and what an institutional venue has to do to be structurally outside that cascade.

N-01-A

The numbers.

tvl · timeline
$99.5BDeFi TVL · 20 Apr 2026 openDefiLlama
$86.3BDeFi TVL · 22 Apr 2026 closeDefiLlama
−$13.2BNet drain · 48 hoursCoinDesk, 20 Apr 2026
~13%Percentage of total DeFi TVLIn two days

Primary source: CoinDesk, 20 Apr 2026.

N-01-B

The chain reaction.

five steps

The exploit itself was narrow. The cascade was not. Five sequential events compressed into roughly fifty hours:

1 · The initial exploit
An attacker took advantage of a flaw in KelpDAO's restaking accounting to mint receipt tokens against collateral the protocol did not actually control. The technical detail of the bug is incidental; the architectural detail is that the receipt token kept its on-chain price reference for hours after it had stopped being backed.
2 · The de-peg
Once the inflated receipt-token supply hit secondary markets, the on-chain price drifted from the implied redemption value. The de-peg was small at first — under 1% — which is the most dangerous size: large enough to be real, small enough to look like ordinary microstructure noise.
3 · The collateral re-rate
Every lending venue that had accepted the Kelp receipt token as collateral suddenly held positions backed by an asset whose price was uncertain. Liquidations began on the most aggressive venues; lenders started pulling deposits from venues that had not yet liquidated.
4 · The deposit run
Adjacent stablecoin and money-market pools, holding no direct exposure to the Kelp token, saw deposits leave too — because the market did not know which book held what. Withdrawals propagated through the lending stack the same way a bank run propagates through deposit accounts.
5 · The TVL drain
By the close of 22 April, total value locked across DeFi had fallen from $99.5B to $86.3B. The exploit's direct dollar impact was a small fraction of that number; the contagion accounted for the rest.
N-01-C

The architectural lesson.

shared = correlated

Restaking is an elegant idea: re-use staked capital to provide security to additional protocols. Architecturally, however, it creates a single asset whose risk is correlated to every place it is accepted. A receipt token cannot be both the source of efficiency and the source of isolation. The protocols that accepted Kelp's receipt token as collateral were not careless; they were doing what the standard composability playbook of the first DeFi wave said to do.

Shared liquidity is shared contagion. The architecture had to change before the next exploit, not after it. imyo · post-mortem conclusion

The fix is not "audit harder" or "shrink the receipt-token universe". The fix is to refuse the architecture that made the cascade possible in the first place.

N-01-D

What contagion-proof actually requires.

silos · oracles · attestation

Three architectural commitments are non-optional for a venue that intends to hold institutional capital:

  • Siloed liquidity by chain. No cross-silo collateral; no synthetic representation of one silo's positions inside the other. A failure on Sui cannot drain Ethereum. A failure on Ethereum cannot drain Sui. See imyo architecture.
  • Restaking-receipt tokens excluded from Tier-1 collateral. Receipt tokens whose risk profile depends on the solvency of an external restaking layer do not belong in the inner book. They can be in capacity-limited Tier-3 with conservative LTVs and explicit risk caps; they cannot be cash equivalents.
  • Oracle dispersion as a pause trigger. When 10+ independent feeds disagree by more than the per-asset tolerance, liquidations pause on that asset until the spread re-converges. The Sentinel Swarm does this automatically and emits an attestation each time it acts. See Sentinel.

None of these is novel as a recommendation. What is novel is putting them into the protocol design from day one rather than retro-fitting them after the next incident. ımyo is the result of doing exactly that.

N-01-E

Read next.

deeper