In April 2026, thieves pulled more than $634 million out of crypto platforms. That's the worst single month since the Bybit catastrophe of February 2025, when losses ballooned to around $1.4 billion, per DefiLlama's tallies. The numbers had been trending down for a while. Then they weren't.

Mitchell Amador, who runs the bug bounty outfit Immunefi, has a theory about why. Speaking at the WAIB Summit in Monaco, he told Cointelegraph that the new wave of frontier AI models is doing most of the damage. He reached for a phrase that's already making the rounds: a "vulnerability apocalypse." Dramatic, sure. But the monthly chart doesn't exactly argue with him.

The machines learned to read code, fast

Amador's argument is simple enough. The current frontier crop (think the latest Claude Opus and ChatGPT releases) handles work that used to eat days of a skilled human auditor's time: reading a smart contract, spotting the edge case nobody tested, drafting the exploit path. The problem, in his telling, is that these tools landed in attackers' hands before defenders worked out how to point them the other way.

That asymmetry matters. An auditor is one person reviewing one codebase on a deadline. An attacker with a capable model can scan hundreds of deployed contracts looking for the same flaw, over and over, at almost no cost. The economics tilt toward whoever moves first, and right now that's the people stealing things.

Worth noting: none of this is new to security as a discipline. Offense usually outpaces defense for a stretch after any capability jump. What's different here is the size of the prize. A bug in a bank's web portal might leak data. A bug in a DeFi bridge leaks the whole vault, instantly, with no chargebacks.

A model that made everyone nervous

The timing of Amador's remarks is hard to ignore. They came just after Anthropic shipped its latest Claude Mythos model, Fable 5, which set off a round of hand-wringing about whether the thing could speed up crypto exploits even further.

Anthropic moved to head that off. The company said on Tuesday that Fable 5 ships with guardrails that punt cybersecurity questions over to a separate model, Claude Opus 4.8. Put plainly: ask Fable 5 something that smells like an attack plan and it hands you off to a different system entirely.

Whether that kind of routing actually stops a determined attacker is an open question. People who want to break things tend to be patient, and creative with how they phrase a request. Safeguards that work in a demo don't always hold up against someone with a financial motive and a weekend to spare. I'd file this one under "reasonable effort, unproven in the wild."

The survival window

Here's the part of Amador's pitch that's actually a forecast rather than a diagnosis. He thinks the crypto industry faces a rough three-to-four-year stretch before defenders catch up. The goal, as he frames it, is for security teams to turn the same models against the problem and start building codebases that attackers simply can't crack.

"Impregnable" was the word he used. That's a high bar, and most engineers would tell you no codebase is ever truly impregnable. But the directional point stands. The defensive use of AI lags the offensive use, and closing that gap is the whole game.

Amador also offered a shortcut. He reckons the window could compress to under two years if more projects embrace what he called crowdsourced security: bug bounties and distributed researcher networks rather than a single audit before launch. Convenient framing for the CEO of a bug bounty platform, naturally. That doesn't make him wrong. The same model that lets one attacker scale up also lets a thousand whitehat researchers scale up, assuming the incentives exist to pull them in.

What the hacks actually looked like

The narrative is tidy. It's worth grounding it in a specific failure, though, because not every exploit traces neatly back to a chatbot.

Take the Kelp DAO hit. On April 19, an attacker drained roughly 116,500 restaked Ether (the rsETH token) from Kelp DAO's bridge. At the time that haul was worth somewhere between $290 million and $293 million. The bridge ran on LayerZero infrastructure.

LayerZero's read on what went wrong is technical but damning. Kelp DAO had configured its setup with a single decentralized verifier network, a 1/1 arrangement, which meant cross-chain messages depended on one verification path. One path means one point of failure. LayerZero said it had warned against exactly that configuration beforehand. Recovery hopes have since faded, with reports that the hacker laundered nearly all of the stolen funds.

Notice what that story is and isn't. It's a design choice that ignored a known warning. You don't need a frontier model to exploit a single-verifier bridge once you've found it. You need to find it. And that's the piece where AI changes the math. Locating the weak configuration across thousands of deployed protocols is the slow, expensive part of an attacker's job. Automate the search, and a flaw that might have gone unnoticed for a year gets found in an afternoon.

So the honest version sits somewhere in the middle. AI didn't create the Kelp DAO vulnerability; human configuration did. But AI plausibly shortens the distance between a vulnerability existing and a vulnerability getting drained, and that's enough to bend the monthly loss chart upward.

What to watch

A few things are worth keeping an eye on over the next several months.

First, whether April was a peak or a floor. One bad month is a data point. Three in a row would confirm Amador's thesis in a way no summit quote can.

Second, how the model providers handle the routing question. Anthropic's approach with Fable 5, shunting security topics to another system, is the first real test of whether labs can meaningfully gate this kind of capability. If researchers demonstrate easy workarounds, expect louder calls for the providers to do more, and louder pushback from people who think gating these tools just hands the advantage to whoever ignores the rules.

Third, the defensive side. If Amador is right that the same models can eventually build near-unbreakable code, somebody has to show it working at scale. Right now that's a promise, not a product. The protocols that survive the next stretch will probably be the ones treating continuous, crowdsourced review as a normal operating cost rather than a launch-week formality.

The uncomfortable truth running under all of this: the tools getting better at defending crypto and the tools getting better at draining it are, in many cases, the same tools. Whoever points them faster wins the round. For now, on the evidence of that April number, the attackers are pointing faster.