The human factor and AI attacks: why code no longer guarantees security in Web3
July 2026 became a landmark month for the decentralized finance industry. The governance code of BonkDAO — the organization behind the Solana memecoin — worked flawlessly, without a single error. However, this did not prevent attackers from draining approximately $20 million from the treasury. There was no hack in the classic sense: the vote was legitimate, and the transaction was executed in full compliance with the protocol's rules.
This incident is not an anomaly but a natural outcome of the evolution of threats. We are witnessing a fundamental shift: hackers no longer break code; they break processes and people. Governance attacks are becoming mainstream, and the human factor is the primary vector for attackers. At the same time, a new threat is already looming on the horizon — AI capable of analyzing vulnerabilities in old contracts faster than any team of auditors.
Attack on decentralization
The attack scheme on BonkDAO is frightening in its simplicity and effectiveness. Disguised as a routine proposal, BIP #76, regarding participant rewards, the hacker embedded a malicious function allowing funds to be transferred to their address. Only $4.4 million was needed to reach quorum — enough to buy 1% of the BONK supply and control nearly 100% of the votes amid extremely low turnout. Six days later, the proposal was automatically executed.
The root of the problem lies in the DAO's architecture itself. The lack of a time delay for execution, a low quorum threshold, and the absence of veto power for the team turned governance into a "hole" in security. This is not the exploitation of a bug in the code but a systemic flaw in decentralization mechanisms. A year earlier, a similar attempt was made against Compound, but it was successfully repelled. In early 2026, Vitalik Buterin had already criticized token-based DAOs, calling them "weak" and "vulnerable to capture." Unfortunately, reality confirms his words.
Audit is not an indicator
While some look for holes in governance, others suffer from technical bugs that, seemingly, everyone has known about for a long time. According to analysts, 90% of smart contracts hacked in 2025 had successfully passed audits. Attackers are not reinventing the wheel but exploiting classic errors: reentrancy, access control issues, oracle manipulation, and logical errors in business logic.
Cross-chain bridges are especially vulnerable. Since 2022, they have lost over $2.8 billion — about 40% of all stolen funds in Web3. The complexity of such systems creates a huge attack surface. The cases of 2026 are telling: the Truebit hack for $26.4 million due to an integer overflow in a five-year-old contract, an attack on the IoTeX bridge validator with $4.4 million withdrawn, and, of course, the largest exploit of the year — the attack on Kelp, where $292 million in rsETH was drained from the LayerZero bridge.
Prompt instead of auditor
A new and perhaps the most alarming trend is the use of AI to find vulnerabilities. In December 2025, Anthropic announced the creation of the SCONE-bench benchmark, consisting of 405 real hacked contracts. Neural networks successfully "hacked" 207 of them, extracting $550.1 million in simulation. Already now, Chainalysis is recording multiple simultaneous attempts to hack dozens of contracts, indicating automated vulnerability scanning. The barrier to entry for hacking is dropping faster than the quality of protection is improving.
The key matters more than the contract
However, no audit would have saved against the largest thefts of recent years. Compromised private keys are the cause of 40% of all stolen funds. The Bybit case, with a loss of $1.46 billion in ETH, is a striking example. Attackers replaced data in the Safe interface, and signers approved a "correct" transaction, handing over control of the wallet. A similar fate befell Humanity Protocol, where keys were stolen via a developer's infected computer. Neither MPC wallets, nor social recovery, nor passkey authentication would have been superfluous in these scenarios.
The industry is moving toward dynamic real-time threat monitoring. Systems like Forta Network and Hypernative have already proven their effectiveness, preventing an attack on Venus Protocol in September 2025. But the main conclusion remains unchanged: perfectly written code is powerless if a person clicks a phishing link or sleeps through a DAO takeover. Perhaps the systemic vulnerability of the blockchain economy lies not in algorithms, but in the fact that there are still people on both ends of the wire.
My view: we are on the threshold of an era where "security through code" is finally giving way to "security through processes." Projects that do not reconsider their governance models and fail to implement preventive monitoring systems will pay a very high price for this lesson. AI hacking is not the future; it is already the present, and ignoring it means signing your own sentence.