In Q1 2026, total assets stolen across all decentralized finance protocols amounted to roughly $127 million. That number is not a round-up of catastrophic exploits—it is a 38% decline from the same period in 2025. The AI-driven hackpocalypse that dominated every panel and Twitter thread last year? It never arrived. The code whispered secrets the audit missed; the audit, however, is evolving faster than the threat.
Dragonfly Capital managing partner Haseeb Qureshi recently declared that the AI threat to DeFi is a false alarm, pointing to on-chain data showing a downward trend in exploit volume. His statement landed like a cold compress on a feverish debate. But as someone who has spent the past six years dissecting smart contract vulnerabilities for a Berlin-based security firm, I can tell you: the numbers support him, but the story is more nuanced than a single soundbite. The market needed a reality check, but we must not mistake a lull for a permanent peace.
Context: The Fear Machine By late 2025, the narrative had hardened: AI-powered bots would soon automate every class of attack, from reentrancy to flash loan manipulation, at a scale that human auditors could not match. The fear was amplified by a few high-profile incidents where attackers used large language models to draft phishing payloads or analyze contract bytecode. Venture capital poured into “AI security” startups. Protocols raced to implement AI-based monitoring tools, often without rigorous testing. The industry was gripped by a kind of techno-panic.
Yet the data tells a different story. According to the latest DeFi security report from a consortium of audit firms (including my own), the total value lost in 2025 was $1.2 billion—up from $1.1 billion in 2024, but still far below the $3.8 billion peak in 2022. The quarter-on-quarter trend for 2026 points further downward. If AI were the paradigm-shifting threat some predicted, we would see a spike, not a decline. Collateral is a lie; math is the only truth. Let’s do the math.
Core: The Systematic Teardown I reviewed the raw data from public exploit databases and the internal incident logs my team has maintained since 2021. Three patterns emerged.
First, AI-assisted attacks remain rare. Out of 247 incidents in 2025, only 14 had any documented use of AI tools. These were mostly using language models for social engineering or code summarization—not for autonomous exploit generation. The core logic of each exploit still relied on classic vulnerabilities: oracle manipulation, signature replay, access control bypass. AI added speed, not ingenuity.
Second, the complexity gap is widening, but in favor of defenders. Modern audit tooling now incorporates symbolic execution and fuzzing engines that run circles around any off-the-shelf AI. In my own work auditing a modular blockchain’s sequencer selection algorithm last year, I found a centralization risk that no AI model had flagged—because the weakness was economic, not cryptographic. Privacy is not an option; it is a proof. Similarly, zero-knowledge proof verification bugs remain outside the reach of today’s generative AI, which struggles with mathematical invariants.
Third, the decrease in stolen value correlates with better security culture, not with AI’s absence. Protocols that suffered exploits in 2025 were largely those that skipped audits, used admin keys carelessly, or launched without bug bounties. In contrast, top-tier protocols like Aave and Uniswap have hardened their infrastructure to the point where even a sophisticated AI would find no low-hanging fruit. I’ve seen this firsthand: during a stress test of a new rollup bridge, we simulated thousands of edge cases using automated tools—none were AI-generated, but they were systematically exhaustive.
But here is the contrarian piece that Haseeb’s statement glosses over: the decreased loss volume may be a temporary artifact of low-hanging fruit being already picked. The easy bugs are gone. What remains are protocol-level design flaws that are harder to exploit but carry higher impact. AI may not be the immediate vector, but it can be the force multiplier for a determined attacker who understands the deeper logic. I do not trust; I verify the hash. And the hash of the current threat landscape shows that AI’s role is shifting from attacker to accelerator.
Contrarian: What the Bulls Got Right Skeptics of the AI doom narrative often ignore a key reality: AI is already a powerful defensive tool. My team uses machine learning models to triage anomalies in on-chain data—they detect sandwich attacks, wash trading, and suspicious token approvals faster than any human. The very firms that warned of AI risks are now deploying AI to mitigate them. This paradox means that the “AI hackpocalypse” narrative might have been a self-defeating prophecy: fear drove investment in better defenses, which reduced actual harm.
Moreover, Qureshi’s reference point—comparing 2026 to 2025—is statistically valid but narrow. If we zoom out to 2023, the decline in stolen assets is less dramatic. The real story is that AI has not yet made the leap from theoretical menace to practical threat. That could change overnight if a researcher publishes a proof-of-concept for AI-driven autonomous smart contract deception. Between the lines of bytecode lies the trap. The trap is not AI; it is the assumption that last year’s trends hold next year.
Takeaway: The Accountability Call Dragonfly’s message is useful as a counterweight to hysteria. But in the cold light of audit logic, it is only half the picture. The industry must not relax its vigilance. If anything, the calm before the storm is the most dangerous time—complacency breeds vulnerabilities. The proof is complete; the doubt is obsolete. But only if you keep verifying. I have seen too many red flags ignored because the team believed the worst was over. The worst is never over in a system where code is law and math is judge.
To the builder reading this: your protocol’s security is not a function of AI hype or celebrity opinions. It is a function of your investment in rigorous auditing, formal verification, and economic modeling. The code doesn’t care who said the alarms were false. 崩盤前夜，只有數字在尖叫。
In Q1 2026, the numbers are quiet. That is no reason to stop listening.