Crypto news

In July 2026, the code of BonkDAO — the governing body of the Solana memecoin — worked flawlessly. Despite this, approximately $20 million were drained from the treasury. No one hacked anything in the traditional sense: the vote was real.

In this piece, I examine how governance attacks are becoming the norm, why the human factor remains hackers' number one choice, and how AI analysis of old smart contracts is turning into a plague for cybersecurity and, at the same time, a master key for attackers.

Attack on Decentralization

On July 6, 2026, BonkDAO faced a takeover of its decentralized governance, resulting in the theft of approximately 4.43 trillion BONK tokens from the treasury. The attacker exploited fundamental weaknesses in the DAO's architecture and community passivity, executing a transaction that was legitimate from the code's perspective.

The attack scheme was as follows: disguised as a routine initiative, Bonk Improvement Proposal #76, which supposedly rewarded voting participants, the hacker added a malicious function that provided for transferring funds from the treasury to his address.

To secure a successful vote, the attacker spent about $4.4 million on centralized exchanges, buying up one percent of the total BONK supply needed to reach quorum. Against a backdrop of extremely low turnout, these funds were enough to control nearly 100% of the votes. After six days, the proposal executed automatically.

The successful attack exposed critical vulnerabilities in the DAO's governance system. There was no time-delay mechanism for executing decisions that would have given developers time to cancel the anomalous transaction. Additionally, the quorum for disposing of colossal sums was too low, and the team had no emergency veto rights.

The attacker read the organization's charter as carefully as any conscientious participant and found a weak spot in it.

This is not the first such case. A year earlier, a similar attempt was made against the Compound protocol's DAO. In July 2024, attackers tried to withdraw 499,000 COMP tokens (about $24.1 million at the exchange rate at the time of the incident) from the protocol's treasury using a malicious improvement proposal. With the organization's typical turnout of 4–5% of the total token supply, such a package proved sufficient to gain control over the vote. Unlike the Bonk case, that attack was spotted and repelled, but only after the proposal was adopted — through negotiations.

Low voting turnout and the concentration of tokens in a few hands are not exploitation of a code bug but a problem with decentralization mechanisms in Web3.

In early 2026, Ethereum co-founder Vitalik Buterin called on the crypto community to develop more advanced and effective forms of DAOs. The programmer called the dominant token-based governance model unproductive and weak. According to him, the current design works but is inefficient, vulnerable to capture, and "completely fails at the task of mitigating the shortcomings of human politics." Buterin saw this as the reason for the community's skeptical attitude toward the concept of such organizations.

Is an Audit Not an Indicator?

In 2025, the number of incidents in Web3 fell by almost half — to roughly 200, compared to 410 a year earlier. At the same time, total losses grew from $2 billion in 2024 to $2.9 billion in 2025. According to Chainalysis experts, the damage from hacks was higher — $3.4 billion. They calculated that three cases accounted for 69% of all losses — for the first time in history, the gap between the largest attack and the average was 1,000-fold.

While holes are being sought in DAO governance without a single line of exploit code, the problems with code are the opposite: the bugs in it are technical and long-known — they just weren't flagged in advance for some reason.

According to Kerberus analysts, 90% of smart contracts hacked in 2025 had successfully passed audits. Most successful attacks did not require finding complex zero-day vulnerabilities — attackers used methods known to the industry for years:

• reentrancy errors — launching a function again before the previous call completes;
• access control errors — when a security-critical function is accessible to a wider circle of people than it should be;
• oracle manipulation — distorting the price data of an asset that the smart contract relies on;
• logical errors in the protocol's business system itself that scanners do not detect.

Cross-chain bridge protocols need quality code reviews more than anyone else. Since 2022, bridges have collectively lost over $2.8 billion — about 40% of all stolen funds in Web3. The reason is that the complexity of cross-chain messaging creates a broad attack surface that specialists physically struggle to assess in its entirety: a large number of independent components must work in parallel, and it only takes an error in one of them.

2026 demonstrated several telling examples with different attack vectors:

1. On January 8, the Ethereum verification protocol Truebit lost $26.4 million due to an integer overflow in the code. The attack was carried out on an outdated TRU token purchase/issuance smart contract deployed about five years earlier. A classic logical error that could have been detected if the audit had been redone after the update.
2. On February 21, attackers gained control over the private key of the IoTeX bridge validator, updated the contract to a malicious version bypassing signature verification, and withdrew $4.4 million, while also minting over 111 million CIOTX and CCS tokens.
3. On April 18, the largest DeFi exploit of the year was the attack on the liquid restaking protocol Kelp: approximately 116,500 rsETH worth $292 million were drained from the LayerZero bridge. The incident became possible due to a configuration with a single validator for cross-chain messages and an attack on LayerZero's RPC infrastructure.

Badges from cybersecurity organizations like CertiK and Code4rena failed to protect the assets of an industry that is still quite young.

Prompt Instead of Auditor

While specialists were studying recurring vulnerability patterns, they gained an adversary for whom such analysis is a routine task, solved at a speed unattainable for any team of human experts.

On December 1, 2025, Anthropic announced the use of AI models to find vulnerabilities in smart contracts and discovered new exploits totaling $4.6 million. To test this, programmers assembled the SCONE-bench benchmark, consisting of 405 smart contracts actually hacked between 2020 and 2025. During the simulation, the neural networks successfully attacked 207 of them, draining a total of $550.1 million. A significant portion of these may have been accessible to LLMs from training data.

To reduce the risk of "reproduction from memory," Anthropic separately tested 34 smart contracts hacked after March 2025.

In practice, this is already happening. In June, Chainalysis specialists reported detecting anomalous activity on networks — multiple nearly identical attempts to hack different contracts simultaneously, pointing to automated vulnerability scanning rather than the manual work of a single hacker who physically could not synchronously attack dozens of targets.

The barrier to entry into hacking is dropping noticeably faster than the quality of defenses is rising. Just a few years ago, finding an exploitable vulnerability in someone else's contract required an experienced Solidity specialist capable of manually reading the code and understanding its logic. Based on SCONE-bench results, for a significant portion of typical attacks, this task can now be reduced to formulating the right prompt.

The Key Matters More Than the Contract

No smart contract audit could have prevented the largest thefts of the past two years.

Private key theft accounted for about 40% of stolen funds — more than any other single attack vector, including smart contract vulnerabilities. Since 2020, attacks on crypto infrastructure have increasingly targeted not the blockchain directly, but the services and infrastructure around it — exchanges, wallets, and interfaces through which people manage assets.

The Bybit hack is telling in this regard. On February 21, 2025, the exchange lost $1.46 billion in ETH, making it the largest single theft in the history of the crypto industry. The attackers interfered with the process of a routine transfer of funds from a cold wallet to a hot one: in the Safe interface, signers saw substituted data and approved an outwardly correct transaction that in reality handed control of the wallet to cybercriminals.

In June 2026, Humanity Protocol suffered a similar fate: the attackers gained access to keys through a backup stored on a developer's computer infected with malware. Losses amounted to about $31 million.

What Actually Works?

When what gets hacked is not the code but the decision-making process, smart contract audits alone are no longer enough. Therefore, the industry is changing both its tools and its very approach to storing and confirming operations:

• MPC wallets — cryptocurrency vaults using multi-party computation technology. Instead of one whole secret key, they create several independent parts stored separately from each other. This eliminates a single point of failure and simplifies access recovery;
• account abstraction with a social recovery mechanism — losing or having a key stolen no longer means the irreversible loss of funds;
• passkey authentication — a passwordless method of logging into accounts that uses cryptographic key pairs and identity verification via biometrics (fingerprint or face scan) or a device PIN. A countermeasure against phishing;
• the use of hardware and cold crypto wallets where software solutions on employees' work machines were previously relied upon.

None of these mechanisms would have been superfluous in the Bybit and Humanity Protocol cases — both incidents were built precisely on the compromise of software and work devices.

Protection is gradually shifting from one-off checks to dynamic real-time threat monitoring. Projects are implementing automated systems like Forta Network and Hypernative, which track anomalous activity in the mempool before a malicious transaction is even included in a block.

The preventive response model has already proven its effectiveness in practice. On September 2, 2025, when attackers attempted to attack the Venus lending protocol, Chainalysis's Hexagate monitoring system detected the deployment of a suspicious smart contract and anomalous activity 18 hours before the main phase of the attack. Twenty minutes before the final transaction that was supposed to drain the pools, the protocol automatically paused. The funds remained safe, and the hackers wasted resources on gas fees.

Web3 architecture was originally designed with a radical goal — to completely eliminate trust in humans by handing control over to immutable mathematics. The hack statistics of recent years demonstrate the opposite effect. Perfectly written and double-checked code turns out to be utterly powerless if a developer clicks a phishing link, and a decentralized community sleeps during a DAO takeover.

My conclusion: We are witnessing a fundamental paradigm shift. An industry built on the principle of "don't trust, verify" faces an irony of fate: the weakest link is the human making decisions. Until DAOs implement mechanisms to protect against majority apathy, and companies move key management to hardware solutions with multi-factor authentication, any technical innovations in security will merely delay the inevitable. Perhaps the main systemic vulnerability of the blockchain economy lies not in the algorithms at all, but in the fact that there are still people at both ends of the wire?