Blockaid's H1 2026 crypto security report delivered a headline that shook every DeFi boardroom: $1.1 billion drained across 212 security incidents. But here's the metric the headline buried. Roughly $790 million — nearly three-quarters of the total — came from privileged key abuse. Not smart contract exploits. Not protocol logic bugs. Keys. Administrative master-switches granting teams the power to upgrade contracts and move treasury funds.
Tracing the ghost liquidity behind the rug pull has been my trade since 2017. This pattern is unmistakable: in 2026, attackers aren't breaking code. They're taking custody.
The report, published by on-chain security firm Blockaid, paints a grim half-year ledger. The marquee losses read like a who's-who of protocol primitives. KelpDAO, a leading restaking protocol, lost $292 million in ether reserves after attackers forged cross-chain messages to bypass its verification layer. Drift Protocol, Solana's flagship perpetuals DEX, lost $285 million in a 12-minute window that left its trading engine powerless. Resolv, a stablecoin issuer, was hit for $80 million through an uncollateralized minting exploit that conjured tokens from thin air. Taiko, CoW Swap, and Verus round out a victim list spanning L2s, DEX aggregators, and cross-chain infrastructure.
Then comes the attribution layer, which is itself a geopolitical story. North Korean state-linked actors are blamed for 55% of total losses — approximately $609 million. This is not a rounding error. It's a sovereign-scale bounty program. And the frequency curve is accelerating: January recorded 18 incidents; June logged 57. If that trajectory holds, the back half of 2026 will not match the first — it will obliterate it.
The security community calls this "the era of infrastructure attacks." Based on my audit work during the ICO boom, when I manually reviewed the Zilliqa Genesis Block contracts, I'd phrase it differently. This is the era of operational security failure. The code was always the easy part to audit. People were the hard part. In 2026, the people won.
Privileged keys are the new vulnerability class
Let me be precise about that $790 million. Blockaid scopes "privileged key abuse" to cases where administrators' private keys were compromised — via phishing, social engineering, or insider access — then used to execute unauthorized functions. KelpDAO alone accounts for $292 million of this. Drift accounts for $285 million. Two incidents, nearly $580 million, both stemming from administrative key compromise rather than contract code execution.
This is a fundamental reclassification of the threat model. During DeFi Summer in 2020, my Python scripts tracked more than 500 Uniswap V2 liquidity pools. The dominant exploit vectors then were price oracle manipulation and contract-level reentrancy. Code flaws. You could audit for them. You could patch them. But key compromise is a human problem. It will not be solved by better compilers or deeper fuzzing. The industry spent four years building better locks, and attackers simply changed their point of entry. They stopped trying to break the lock and started stealing the key.
Cross-chain verification is the skeleton in the closet
The report lists KelpDAO, Taiko, and Verus as cross-chain bridge victims. Attackers bypassed verification through "forged proofs and assertions." A cryptographic proof system cannot be forged without breaking the underlying assumption. If attackers forged proofs, the system was not verifying proofs on-chain in the first place. It was relying on off-chain multi-sig signers, relayers, or sequencers for coarse-grained confirmation. I call this pseudo-decentralized validation.
The same pattern appears across the ecosystem's most fragile layer. When I audited the Zilliqa sharding contracts in 2017, the overflow bug was fixable because the system was code-native. A bridge that depends on a relayer's say-so has a human in the loop. And a human in the loop is a human with a price.
Verus had the same bridge contract exploited twice. That is not a technical failure. It's an organizational failure — the inability to sunset a compromised deployment. The code doesn't lie; it does exactly what it's told. The problem is who is doing the telling. Decentralized sequencing has been a PowerPoint slide for two years; the actual sequencing, relaying, and signing remains in the hands of a few operators. Attackers know it. That's why they don't target the cryptography. They target the people holding the pen.
New attack surfaces: EIP-7702 and the AI agent
Two novel categories appeared in H1 2026. Four incidents involved EIP-7702 wallet delegation attacks — the post-upgrade feature allowing wallets to delegate transaction execution to smart contract logic. The upgrade was designed for better UX and smart-contract wallets on existing EOAs. It also created a privilege delegation path that attackers exploited at least four times to redirect funds. The standard is new. The security hardening hasn't caught up.
Then there's Bankr, an AI-agent-powered protocol that lost $216,000 to prompt injection. Let me be blunt about the dollar size: that's a rounding error in this report. As a structural signal, it's terrifying. An AI agent executed a malicious instruction because the prompt was weaponized. The agent didn't verify intent. It didn't check provenance. During my NFT metadata forensics work in 2021, I learned that the provenance was always in the data — the IPFS hashes, the contract records, the metadata links. Metadata holds the provenance the price ignored. Here, the metadata wasn't even consulted. This is the fusion of social engineering, which tricks humans, and prompt injection, which tricks machines. The AI layer inherited all of our frailties, with none of our paranoia.
Legacy contracts refuse to die
The report also flags old smart contracts still causing losses — Aztec Connect and Raydium's AMM V3 among them. These contracts remained live for years after their projects moved on. In a bull market, there's no economic incentive to retire old code — only the cost of migration. Following the exit liquidity to its cold storage, I've noticed a recurring rhythm: abandoned contracts are treasure maps for careful adversaries. The contract doesn't have to be new to be profitable. It just has to hold assets.
Correlation is not causation
Before concluding that bridges and keys are dangerous, consider what the report doesn't say. Security firms produce these audits to establish a market for their services. Blockaid's report is simultaneously a data disclosure and a product pitch. That doesn't invalidate the numbers, but it incentivizes a specific framing: incidents are rising, the industry is fundamentally unsafe, and you need enterprise-grade monitoring to sleep at night.
There's also a measurement artifact. Detection tooling improved dramatically between 2024 and 2026. Incidents that would have been missed or misattributed are now caught and catalogued. The rising count is partly a better-net story. Whether actual risk per transaction dollar increased is a question this report doesn't answer.
The attribution data carries its own policy risk. North Korea at 55% will trigger regulatory scrutiny. Governments will respond with tighter compliance mandates, harsher KYC expectations, and more aggressive address sanctioning. The systemic risk nobody quantifies is regulatory overreaction to a problem that is, at its core, a key-management problem. The panic response to hacks may end up being the most expensive attack of all.
What I'm watching next
Protocol treasuries now face a key-custody dilemma: holding keys invites attack; distributing custody invites governance paralysis. The market will begin pricing security provenance — the documented history of a protocol's key custody, incident response, and bridge validation design — into token valuation. Chasing the gas fees through the mempool labyrinth this week, I'm checking who is implementing threshold signatures and hardware-backed key isolation in response to this report, and who is merely issuing a PR statement. The protocols building at the physical security layer, not the marketing layer, will capture the next capital cycle.