ADVERTISEMENT

Events

AI Can Scan a Smart Contract for $1.22, but 2026’s Biggest Crypto Thefts Began With People and Vendors

In December 2025, Anthropic researchers reported that an AI agent could scan a newly deployed smart contract for exploitable flaws at an average compute cost of $1.22, a price that makes probing public code for weaknesses close to free and that set off months of warnings that AI would soon overwhelm DeFi security.

DeFi Planet reviewed the largest documented crypto thefts of 2026, from Drift’s $285 million loss on April 1 to Binance’s roughly $6.34 billion on October 29, and sorted each by the layer that failed first. The eight incidents total nearly $8 billion in gross losses, and only one began with a flaw in smart contract design, a price manipulation at Tectonic that Cronos validators reversed with a chain rollback, leaving about $9.2 million unrecovered. Roughly 88% of the total came from compromised people, credentials, verifier configurations or tampered supply chains, and about 10% came from software or firmware defects outside smart contracts, at Bitget, Liquid and Coldcard. 

AI did appear in the largest theft of the year, as a lure and a foothold and not as the vulnerability scanner that security teams had braced for. Binance’s initial report says a Wallet DevOps engineer downloaded a trojanized Homebrew installer in June while trying to install an AI coding tool, and that attackers later abused an internal AI tool on October 20 to take control of two more engineers’ machines. 

Cheap AI scanning does change the economics of finding bugs in public code, and anyone who waves that away isn’t paying enough attention. The 2026 losses nevertheless show attackers collecting most of their money through the people, vendors and verifier setups around the code, and a defender who audits only the contracts leaves the largest exposures unexamined. 

A Co-Founder of OpenZeppelin Called All of DeFi Unsafe

One of the loudest alarms came from Manuel Aráoz, a co-founder and former chief technology officer of smart contract security firm OpenZeppelin, who wrote in a May 2026 post on X that he no longer trusted any DeFi protocol. 

“I now consider all of DeFi unsafe. Coding agents are superhuman at finding vulnerabilities, and smart contract security is too asymmetric: defenders need to fix every bug while attackers need just one exploit to steal funds.”

Aráoz said he had told friends and family to close their DeFi positions, including on long-running protocols such as Aave and Compound. OpenZeppelin responded that his claims “do not represent the company’s position” and called the AI threat “real but manageable,” saying the answer is better security and not an exit from DeFi. 

The warning carries weight because it’s coming from a founder of a firm whose contract libraries and audits are widely used across DeFi, and the economics he describes are real, since a defender must close every flaw while an attacker needs one. 

In April, Anthropic announced Project Glasswing, a restricted program that gives selected organizations access to Claude Mythos Preview, a model the company said had found thousands of serious vulnerabilities in major operating systems and browsers. April also became the costliest month for crypto theft up to that point in 2026, at about $642 million, according to Global Ledger. The two events were unrelated in cause, because the April losses came from Drift’s signers and Kelp’s bridge verifier and not from model-found bugs, but together they explain why warnings of an AI-driven wave of DeFi exploits sounded credible by summer. 

Put those events together, and the “AI is about to eat DeFi” starts to sound less like a hot take and more like a forecast. The actual problem is yet to unfold.

Why Cheap AI Scanning Changes the Economics

Let us give the worried crowd their due first, because the economics are real and true.

A professional smart contract audit typically costs between $25,000 and $100,000 for a DeFi protocol and can take several weeks for a mid-sized codebase, according to Sherlock’s 2026 pricing reference. That high cost provides the benchmark against which an attacker’s AI scan is evaluated.

The researchers ran AI agents against 2,849 newly deployed BNB Chain contracts with no known vulnerabilities. Scanning all of them with GPT-5 cost $3,476, an average of $1.22 per contract, and the agents found two previously unknown vulnerabilities. 

The $1.22 covers the cost of a scan and not the cost of a finding. Dividing the $3,476 bill by the two vulnerabilities gives $1,738 per bug, and because the average simulated exploit yielded $1,847, the net profit was about $109 per bug, in a simulated environment and not on a live network. 

Even so, $1,738 per bug is a small fraction of one audit fee, and public bytecode that is hard to change after deployment lets an attacker sweep thousands of contracts at a fixed cost per scan. A sweep is still a numbers game, since the study’s two hits came from 2,849 contracts, about 0.07%, so the attacker’s margin depends on how much money sits in the flawed contracts and not on the scan price. That asymmetry is the one Aráoz described, and on that point he is right.

Aráoz’s argument leaves out what happens after a bug is found, because stolen money requires both funds sitting in the vulnerable code and a route to cash out an exploit that works. 

What Actually Failed in 2026’s Eight Biggest Thefts

To test whether AI scanning explains where the money went, DeFi Planet compiled the eight largest documented crypto thefts of 2026 and sorted each by the layer that failed first. 

Table: 2026’s eight largest crypto thefts, by the layer that failed first.

Date (2026)VictimStolen (approx.)What failed firstOn-chain protocol flaw?AI’s public role
29 OctBinance (exchange)$6.34BEngineers’ computers compromised through a trojanized installer and a fake job interview; attacker reached an internal wallet signing serviceNoBait (fake AI-tool installer) and an abused internal AI tool
24 SepBitget (exchange)$387.5MZero-day in a third-party security appliance, then a jump to the wallet job serverNoNone disclosed
6 SepLiquid Network (Bitcoin sidechain)$320MRange-proof cache bug in Elements node softwareNo (node software)Bug reproduced by AI after the attack
18 AprKelp DAO (liquid restaking)$292MPoisoned RPC nodes fooled a single-verifier bridge setupNoNone disclosed
1 AprDrift (Solana trading)$285MSecurity Council signers socially engineered into pre-signing transactionsNoNone disclosed
30 AugTectonic (Cronos lending)$120.4M affected; ~$9.2M escapedPrice of a thinly traded collateral token pumped; the price feed followedYes (protocol design)None disclosed
30 JulColdcard users (hardware wallet)$116MFirmware build error made some devices generate weak seedsNoNone disclosed
OctLedger buyers (via a reseller)$86M–$93MDevices fitted with seed-stealing implants and resold through the distribution channelNoNone disclosed
Total8 incidents~$7.9BSeven of eight failed outside on-chain protocol code1 of 81 of 8 attacks (Binance)

Our Methodology: DeFi Planet reviewed every crypto theft of roughly $85 million or more reported between January 1 and October 31, 2026, using DefiLlama and CertiK records and each victim’s disclosures. Figures are gross losses before recoveries, and recoveries are listed separately. A $284.8 million incident that CertiK lists against an unidentified victim was excluded because no public account of its cause exists, so the table shows the eight largest documented thefts and not the eight largest overall. Categories reflect the first point of failure and not later laundering, and AI involvement counts only where it is publicly confirmed. 

A few things stand out:

  • Only one of the eight began with a flaw in smart contract design. At Tectonic, Cronos validators rolled back about $111.2 million of the roughly $120.4 million affected, which leaves about $9.2 million lost and puts the incident well below the $85 million threshold on a net basis. It is in the table on gross exposure and accounts for about 1.5% of the total. 
  • Excluding Binance, the other seven total about $1.6 billion, and protocol design accounts for about 7.5% of that. 
  • Binance, Bitget, Kelp and Drift together lost about $7.3 billion, around 92% of the total, and each has been linked to North Korean groups at different levels of confidence. 
  • AI appears in two incidents. At Binance, it served as a lure and an over-trusted internal tool, and at Liquid, AI agents reproduced the bug after the funds had left. Public reporting shows no case where AI found the flaw first. 

READ ALSO: Do KYC Databases Make Crypto Holders Targets of Wrench Attacks?

The Big Thefts Started With People, Not Code Scans

Meir Dolev, co-founder and CTO of blockchain security firm Cyvers, said in May that there is still limited public forensic evidence of AI directly carrying out DeFi exploits. What has been verified, he said, is the broader rise in AI-assisted scams.

Dolev listed smart contract logic among the most exposed areas, alongside admin keys, DevOps pipelines, front ends, signer workflows and social engineering of the people who hold access. His advice was to treat security as a continuous control during execution instead of a pre-launch checkbox, not to leave DeFi. Six months on, that list reads like a preview of the year.

April’s two largest thefts, Drift and Kelp DAO, made up about 90% of the month’s stolen funds. Neither came from a flaw in smart contract code.

Drift lost about $285 million through its signers

Attackers took about $285 million from Drift, a Solana-based trading platform, on 1 April by seizing its Security Council’s admin powers. According to Drift’s own investigation, people posing as a quantitative trading firm spent six months building relationships with Drift contributors, including at industry conferences. They then obtained approvals on pre-signed transactions using Solana’s durable nonces, a feature that lets a signed transaction be held and executed later. Drift said in June that Mandiant attributed the attack to UNC6862, a North Korea-linked group. No smart contract bug or leaked seed phrase was involved.

This attack succeeded because human signers were manipulated over six months of build-up and networking, not because an AI scanned the contract for bugs.

Kelp DAO lost $292 million when its bridge believed a lie

Seventeen days later, on 18 April, an attacker drained 116,500 rsETH, worth about $292 million, from Kelp DAO’s cross-chain bridge. The bridge relied on LayerZero, a messaging protocol that tells one blockchain what happened on another. According to LayerZero’s post-mortem, attackers compromised two RPC nodes (the servers that feed blockchain data to the verifier), then flooded the healthy nodes with traffic so the system fell back to the poisoned ones. Kelp ran a single-verifier setup, so one deceived verifier was enough to approve a forged message. In effect, the bridge was convinced the thief was the rightful owner of the funds.

LayerZero linked the attack, with preliminary confidence, to North Korea’s Lazarus Group. Kelp pushed back. In a statement on X, it said its one-verifier configuration matched LayerZero’s documented default and that LayerZero had confirmed the setup. “Establishing a shared and accurate account of what happened is the foundation for making the right fixes together,” Kelp’s team wrote. 

DeFi Planet reported how the RPC attack exposed cross-chain verifier risk at the time.

Our take: both sides can be partly right. A default that lets one fooled verifier release $292 million is a bad default, and a team holding that much value should have overridden it anyway. Public records show no evidence of AI involvement; the exploit simply relied on controlling two servers and targeting a setup without a secondary verifier.

Neither incident has been conclusively tied to AI. Both hit infrastructure or governance, and both were set up over weeks or months by human operators.

READ MORE: Crypto Security Has Come a Long Way, So Have the People Trying to Break It

Case Study 1: Liquid Network and the Overlooked Sidechain Bug

Liquid is a Bitcoin sidechain run by a federation of companies and developed largely by Blockstream. Users lock bitcoin to receive L-BTC, a 1:1 token that moves faster and more privately on Liquid. Every L-BTC is supposed to have real bitcoin behind it.

On 6 September 2026, an attacker created about 3,998.5 L-BTC with nothing behind it and converted it to bitcoin through Liquid’s peg-out. Roughly 4,000 BTC, worth about $320 million, left the federation’s wallet. Liquid said no keys were compromised and its operators were not hacked.

So what broke

Liquid hides transaction amounts and uses range proofs to show they are valid without revealing them. Checking these proofs is expensive, so Liquid’s node software, Elements, caches results it has already verified.

Blockchain security firm SlowMist traced the problem to how the cache created its label entries. Because several variable-length fields were combined without length prefixes, different inputs could generate identical cache labels. As a result, a forged proof could match a label already marked as valid and pass verification. To resolve this, Blockstream released Elements v23.3.4, which adds length prefixes and allows users to disable the cache entirely.

Why it matters for the AI debate

Liquid is the incident that comes closest to the scenario AI-security warnings describe, because the flaw sat in node software securing a sidechain that held hundreds of millions of dollars, a category that gets far less audit attention than lending contracts. 

Nothing published so far shows that AI found this bug. After the incident, a mempool space contributor known as Orangesurf claimed that Anthropic’s Fable 5 model could spot the flaw from a five-word prompt.

 

Alpen Labs went a step further. Its CEO, Simanta Gautam, said his AI agents traced the cache flaw and reproduced it locally in about an hour.

After the incident, mempool.space contributor Orangesurf claimed that Anthropic’s Claude Fable 5 could spot the flaw from a five-word prompt, and Alpen Labs chief executive Simanta Gautam said his team’s AI agents traced the cache flaw and reproduced it locally in about an hour. Both claims came after the exploit was public, so they show that AI can reproduce a known bug and say nothing about whether it would have found the bug before the attacker did. 

The timeline raises a sharper question. According to ForkLog, a change adding the missing parameters to the cache was merged into the public Elements repository on September 1, five days before the attack, although no release containing the fix had shipped. A merged but unreleased security fix lets anyone, including an automated reviewer, read the patch and work backwards to the flaw, and whether the attacker did so is unknown. For operators securing hundreds of millions of dollars, the gap between merging a fix and shipping it is an exposure window that deserves its own policy. 

The aftermath

The attackers, who called themselves white hats, returned 3,400 BTC on 7 September. They kept about 598.5 BTC, worth roughly $47 million, and left an on-chain message demanding a bounty. Blockstream refused on 11 September, and it didn’t mince words:

“Taking assets without authorization and withholding their return is a crime, not responsible disclosure. It is not white-hat activity. It is theft.”

Liquid resumed block production on 10 September, but direct peg-outs back to Bitcoin stayed restricted for weeks, and in October reserves covered only about 86% of L-BTC in circulation. In early November, Liquid announced that full 1:1 backing had been restored after a final round of deposits, which Liquid Federation board member Samson Mow attributed to Blockstream.

AI Can Scan a Smart Contract for $1.22, but 2026's Biggest Crypto Thefts Began With People and Vendors
The Liquid incident, from the unreleased fix to restored backing. Source: X

Case Study 2: Bitget and the Oldest Weak Point

Eighteen days after Liquid, Bitget got hit through a door that had nothing to do with blockchain code.

The timeline

At 18:31 UTC on 24 September, Bitget detected unauthorized transfers from its hot and warm wallets (wallets kept online so withdrawals can be processed quickly) and paused all withdrawals. Its first estimate was $351.6 million. A day later, after tracing stolen funds on Zcash and Tron, Bitget raised the figure to $387.5 million, spread across seven networks. Withdrawals reopened in phases: Bitcoin on 28 September, Ether on 29 September, and USDT on 30 September.

How the attacker got in

Forensic findings from Mandiant and SlowMist, published by Bitget on 30 September, describe a classic intrusion. According to Mandiant, the attacker gained privileged access to two third-party security appliances on Bitget’s network. From one of them, the attacker moved sideways to Bitget’s production wallet job server and installed malicious packages, according to BleepingComputer, which reviewed the reports.

SlowMist found that the entry point was a zero-day flaw in one vendor product, with the earliest malicious activity in the logs dating to 31 August. That’s more than three weeks of quiet access before any money moved. Think about that the next time someone tells you the threat is a bot that finds bugs in seconds.

Bitget CEO Gracy Chen said the flaw gave the attacker high-level internal credentials, which were used to issue fraudulent withdrawal commands. She was blunt about what didn’t happen: “Private key compromise has been ruled out.” And about what did happen: “The attacker compromised a critical backend system within our wallet infrastructure.” Bitget also said its cold wallets were untouched.

Who did it, and who pays

On the day of the hack, Chen said IP addresses tied to the attack pointed to North Korea, and DeFi Planet reported that XRP trails pointed toward the Lazarus Group. Analytics firm Elliptic later said the laundering pattern resembled earlier North Korean thefts. By 28 September, Chen described the North Korea link as a preliminary indicator still under assessment, and the published forensic findings do not name an attacker. For now, the attribution is Bitget’s early suspicion backed by Elliptic’s analysis, not a confirmed finding.

Bitget said its User Protection Fund, which held more than $464 million at the time, would absorb the full loss, so customer balances stayed intact. The exchange has since added independent verification for withdrawals and tightened internal access.

No AI bug-hunter was needed here either. The attacker needed a vendor zero-day, patience, and access to the system that signs withdrawals.

AI Can Scan a Smart Contract for $1.22, but 2026’s Biggest Crypto Thefts Began With People and Vendors
Bitget hack: the way in was a vendor appliance, not the blockchain. Source: BleepingComputer

Case Study 3: Binance, Where AI Showed Up as the Bait

Then, on 29 October, every number above suddenly looked small.

What happened

Starting at around 20:05 UTC, attackers drained funds from Binance wallets across ten different blockchains. Most of the money was taken within just 20 minutes in what DeFi Planet reported as the largest crypto theft on record. In a November 3 update, Binance confirmed total damages of roughly $6.34 billion, though about $2.85 billion was eventually frozen or intercepted, Security Affairs reported. Blockchain analytics firm TRM Labs attributed the attack with “high confidence” to TraderTraitor, a North Korean cybercrime unit, according to Euronews.

Binance co-CEO Richard Teng’s first explanation, on X, was short: “Our initial investigation suggests this security incident may have started with a compromised DevOps engineer’s laptop.” He later added that this was “not a single ‘one-click’ action”.

How the attacker got in

Binance’s initial findings, published on 3 November, describe a campaign that took months, not minutes:

  • June: an engineer on Binance’s Wallet DevOps team searched online for an AI coding assistant and installed a trojanized Homebrew installer. Binance caught it, replaced the device and rotated credentials, but details about the engineer’s role and colleagues had already leaked.
  • September: a second engineer was approached on a job-hunting platform by someone posing as a “Project Director.” After two rounds of interviews, the engineer was asked to download an “assignment” repository onto a work computer. It carried malicious code that ran during setup.
  • 20 October: from that machine, the attacker abused an internal AI tool to gain control over the first engineer and a third, according to The Hacker News’ reading of Binance’s report.
  • 29 October: the attacker logged into the third engineer’s computer, used a container deployment token stolen from the first engineer, and reached “an internal wallet signing service to initiate transactions outside the normal withdrawal process.”

Binance said there was “no evidence of any wallet private key leak,” and its report is clear about what kind of incident this was: “It was not a single isolated breach or a simple phishing incident.” Binance has said it will cover user losses, and by 31 October it had resumed withdrawals and said 90% of stolen user assets were covered.

Why this is the AI story that matters

The AI in the year’s largest theft was never a model flaw. It was the reason an engineer went looking for a download, and later an internal tool with enough reach that, once an attacker had a foothold, it helped take over two more engineers’ machines. Binance’s report describes no vulnerability in any AI model, and the weak point was what the internal tool was allowed to touch. 

Solana co-founder Anatoly Yakovenko put it more bluntly on X after the hack: “Phishing and social engineering is the biggest threat to crypto. Not AI breaking math or quantum computers.

Dedicated iOS in lockdown mode and @multisig is probably the best low-cost and high roi setup anyone can have.” 

Let us add one more thing: AI tools are now part of the social engineering surface, both as bait and as a privileged insider nobody thinks of as an insider.

Two Fronts of Risk, and a Newer One

When you line up all eight threats, you get two separate risks, plus a newer one growing inside the second. 

Front one: cheap scanning pushes attackers toward under-audited code

If scanning a contract costs about a dollar, attackers can sweep everything. The richest targets won’t always be the famous lending protocols, which get audited over and over. They’ll be the code underneath: bridges, sidechains, federation software, wallet firmware and signer tooling, where a Liquid-style cache bug or a Coldcard-style build error can sit unnoticed for years. That’s where I’d expect AI-assisted discovery to show up first.

Front two: operational security is still the oldest and most expensive failure

This is where the money actually went; Binance, Bitget, Drift, and Kelp together lost about $7.3 billion without a flaw in any of their smart contracts. Social engineering, poisoned servers, vendor products, and hot-wallet signing systems did the damage. Even before Binance, Bitget and Liquid alone made up about 92% of September’s roughly $766 million in hack losses, according to CertiK. Add Binance’s $6.34 billion to CertiK’s $2.68 billion in gross losses through September 2026, and it is already past $9 billion, with about 70% of that from a single attack.

The newer problem: AI tools inside the perimeter

The Binance attack illustrates where operational security threats are heading. Today, exchange and protocol teams use AI assistants and automated tools that have direct access to internal code, credentials, and deployment systems. Every tool introduced creates a potential target for attackers to impersonate and another account with elevated access privileges. If security teams focus solely on auditing smart contract code with AI without auditing what internal AI tools can access, they miss the primary vector of modern attacks.

ALSO READ: The Future of Smart Contract Audits with AI

How Users Can Limit Exposure

  • Prefer protocols with continuous security, not a one-off audit: A report from launch day says little about code that has changed since, or about the people and servers around it. Look for ongoing audits, live monitoring, and an active bug bounty.
  • Check what a bridge or custodian actually trusts: Ask how many independent verifiers must approve a cross-chain message, who controls peg-outs and admin functions, and how upgrades are approved. Kelp’s single-verifier setup was a choice, and it was the failure point.
  • Treat exchange hot-wallet exposure as a risk you can limit: Keep only what you’re actively trading on an exchange. Binance and Bitget both covered their users this time, but a protection fund is a promise, not a guarantee. Check whether your exchange publishes proof of reserves and how large that fund is.
  • Buy hardware wallets directly from the manufacturer and watch firmware advisories: Ledger losses came from devices tampered with in the distribution channel, and Coldcard losses came from weak seeds created on older firmware. Moving long-term holdings to self-custody still makes sense, but as TRM Labs put it after Coldcard, “self-custody relocates risk rather than eliminating it.”

Cheap AI scanning lowers the barrier to discovering code vulnerabilities, but it doesn’t make human signers any less susceptible to manipulation. In 2026, those human and operational security vectors were precisely where the vast majority of stolen funds were lost.

Frequently Asked Questions

Can AI hack smart contracts?

In controlled tests, yes. Anthropic’s December 2025 study showed AI agents finding and exploiting real smart contract flaws in a blockchain simulator, at an average scan cost of $1.22 per contract. There is still limited public evidence that AI has carried out a live DeFi exploit on its own.

How much of 2026’s crypto theft is linked to North Korea?

Elliptic attributes more than $1 billion in 2026 thefts across more than 51 incidents to suspected North Korean operations, about 37% of CertiK’s $2.68 billion gross total through September, and that was before the Binance theft.

What is the difference between gross and net crypto hack losses?

Gross losses count everything taken, while net losses subtract assets later frozen, returned or reversed. CertiK’s $2.68 billion gross figure through September falls to about $2.26 billion after $420.4 million was frozen or returned.

Do smart contract audits prevent crypto hacks?

Audits examine code at a point in time, so they do not cover stolen credentials, compromised vendors or verifier configurations. Most large 2026 losses began outside the contracts an audit reviews.

What happens to users’ funds when an exchange is hacked?

It depends on the exchange’s reserves and protection funds. Bitget said its User Protection Fund, which held more than $464 million, would absorb the full loss, and Binance said it had replenished 98% of impacted assets as of November 3.

Disclaimer: This article is intended solely for informational purposes and should not be considered trading or investment advice. Nothing herein should be construed as financial, legal, or tax advice. Trading or investing in cryptocurrencies carries a considerable risk of financial loss. Always conduct due diligence.

Enjoyed this? Bookmark DeFi Planet, explore related topics, and follow us on Twitter, LinkedIn, Facebook, Instagram, Threads, and CoinMarketCap Community for seamless access to high-quality industry insights.

Take control of your crypto portfolio with DEFI PLANET PRO, DeFi Planet’s suite of analytics tools.

ADVERTISEMENT
ADVERTISEMENT

Spotlight

ETH $1,797.10 -1.16% NU $469.04 +0.00% POLY $792.15 +0.00% APE $748.09 +0.00% FET $226.49 +0.00% ARPA $160.56 +0.00% GTC $2,668.91 +0.00% FORTH $4,194.39 +0.00% PLU $2,124.48 +0.00% MLN $10,448.26 +0.00% ETH $1,797.10 -1.16% NU $469.04 +0.00% POLY $792.15 +0.00% APE $748.09 +0.00% FET $226.49 +0.00% ARPA $160.56 +0.00% GTC $2,668.91 +0.00% FORTH $4,194.39 +0.00% PLU $2,124.48 +0.00% MLN $10,448.26 +0.00%