TechFlow Logo
Login/ Sign up
ETH Gas
Gwei
Fear
gas
Research: Nearly 90% of Stolen Funds Cannot Be Recovered, Attack Targets Shifting from Code to "People"

Research: Nearly 90% of Stolen Funds Cannot Be Recovered, Attack Targets Shifting from Code to "People"

2026.07.31
Share
In an era where AI-synthesized reality becomes possible, what needs to be verified is no longer a single piece of information, but the environment itself.

Author: WuSay Blockchain

In an era where AI-synthesized reality becomes possible, what needs verification is no longer a single piece of information, but the environment itself.

In the first half of 2026, the crypto industry experienced 182 public security incidents, with losses totaling approximately 956 million USD. More alarming than the total loss is the destination of the funds: according to SlowMist statistics, only 18 incidents in the first half resulted in recovered or frozen stolen funds, totaling approximately 118 million USD, accounting for 12.3% of the total loss. The remaining nearly 90% of stolen funds are irrecoverable.

Another figure is easily misinterpreted: the 956 million USD represents a decrease of nearly 60% compared to the same period last year, but this does not mean the industry has become safer. The decline in losses is almost entirely due to the non-recurrence of last year's single massive Bybit incident (approximately 1.5 billion USD); the number of incidents actually increased by about 50% year-on-year. Attacks have not weakened but have changed direction—from targeting protocol contracts to targeting people.

The two largest loss incidents in the first half did not succeed by breaching smart contracts: Drift Protocol lost approximately 285 million USD through a six-month social engineering infiltration, starting with a few transactions signed by a multisig signer that appeared to have "no impact"; a victim in Singapore was invited into a video conference where all government officials were AI-generated, losing approximately 4.9 million SGD. The most expensive vulnerabilities appeared in people.

This is one of the core judgments of the "2026 OKX Web3 Security Half-Year Report" released jointly by the OKX Web3 Security Team, SlowMist, and OtterSec, sharing the same conclusion as SlowMist's concurrently released "2026 First Half Blockchain Security and Anti-Money Laundering Report." The two reports come from a trading platform and a security company respectively, yet point to the same judgment: the frequency of attacks is rising, methods are evolving, and targets are shifting from "code" to "people."

Losses Down Nearly 60% Year-on-Year, Attack Activities Rise Instead

The same set of data, sliced by different metrics, yields opposite conclusions; this is the premise for understanding the security posture of this half-year.

According to OKX's report citing SlowMist's Hacked Archive statistics, there were 182 incidents in the first half with losses of approximately 956 million USD; for the same period in 2025, there were 121 incidents totaling approximately 2.373 billion USD. The number of incidents increased by about 50% year-on-year, while the amount decreased by about 60%. The main reason for the decline in amount is that the February 2025 Bybit incident (approximately 1.5 billion USD, where hackers compromised the computer of a Safe{Wallet} developer for the exchange and tampered with official website scripts) was an extreme case. OKX's report judges that: excluding this outlier, comparable losses this year have risen instead of fallen.

The OKX report further splits attack causes by dual metrics: by number of incidents, contract and logic vulnerabilities remain the top cause, totaling 85 cases; but by loss amount, supply chain attacks rank first at approximately 298 million USD, followed by contract vulnerabilities (approximately 152 million) and private key leaks (approximately 130 million). SlowMist summarizes this characteristic in eight characters: incidents are decentralized, losses are centralized. Among these, truly fatal large losses are concentrating on key links such as infrastructure, cross-chain, and supply chains. An unbalanced footnote is: cross-chain bridges experienced only 20 incidents in the half-year, yet caused approximately 346 million USD in losses.

Funds are no longer primarily lost through contract vulnerabilities. The OKX report summarizes the changes in this half-year into three main lines: large losses are increasingly occurring outside contracts; ordinary users have become the main attack targets; and AI Agents have shifted from tools to prey. Additionally, there are two threads running throughout: supply chain poisoning penetrating the development link, and the money laundering downstream where all stolen funds eventually flow. These are expanded upon below.

Projects: The Largest Losses Occur Where Audits Cannot Reach

Most projects with the most severe losses in the first half were not unaudited; the points of failure were precisely outside the audit scope: signature processes, cloud keys, validation nodes, and developer devices. Audits can prove contract logic is correct, but cannot prove these operational links are secure.

KelpDAO: It was not the contract that was breached, but the validation path. This was the largest single loss in the first half. According to OtterSec's review, attackers polluted LayerZero's internal RPC nodes on one side while launching DDoS attacks on honest external nodes. Cross-chain messages should have been cross-validated by multiple independent nodes, but the bridge当时 adopted a "1-of-1" single validation node configuration. The sole validation point received only forged data, ultimately approving a withdrawal without real asset backing; approximately 116,500 rsETH were transferred out, of which approximately 75 million USD was frozen afterwards. Single validation nodes have long been considered a high-risk configuration; in the past, this was just a theoretical hint in architecture reviews, but after the KelpDAO incident, it became a realistic loss of nearly 300 million USD. SlowMist added details on the destination of the stolen funds: LayerZero attributed the incident to the Lazarus sub-group TraderTraitor; the stolen tokens were subsequently used as collateral to borrow approximately 236 million USD in real assets (WETH) on lending platforms like Aave, once triggering a liquidity crisis in the DeFi market.

Drift: A six-month signature inducement. Its mechanism is worth explaining: durable nonce is equivalent to a transaction voucher "pre-signed and redeemable at any future time." Attackers operated for six months under the guise of a quantitative institution with over 1 million USD in real deposits to build trust, inducing signers to pre-sign such management transactions. At the time of signing, there was no visible impact; when the project subsequently adjusted the multisig threshold, creating a window lacking timelock (time lock: sensitive operations must be publicized and delayed before taking effect) protection, attackers broadcast the pre-obtained signed transactions, completing 31 withdrawals within 12 minutes and making off with over half of the locked assets. The core lesson of this incident is: transactions that have "no impact" at the time of signing do not mean they will not be redeemed in the future; blind signing, pre-signing, and unparseable management transactions should all be treated as high-risk operations.

Portraits of single-point failures. Resolv Labs' AWS cloud keys were stolen, allowing attackers to mint approximately 80 million unpegged tokens; Step Finance suffered from executive devices being compromised, with private keys used to empty the treasury; Humanity Protocol involved developer devices infected with malware and loss of private key control. SlowMist mentioned that on-chain analyst ZachXBT tracked and found mixing behavior between its stolen funds and those from KelpDAO, basically ruling out internal作案 possibility, pointing again to Lazarus.

The SlowMist report also recorded a set of more underlying cases, pushing attacks down to the level of cryptographic engineering implementation. The common point of the three incidents is: vulnerabilities were not within the regular coverage of code audits.

• Taiko (June 22, approximately 1.7 million USD): A signature private key was mistakenly committed to a public GitHub repository; attackers used this to forge Layer 2 network state proofs and trick funds release, causing the TAIKO token to drop over 20% in the short term.

• SecondFi / formerly Yoroi (June 21–23, approximately 2.4 million USD): The signature algorithm implementation lacked a necessary random masking step, allowing a complete private key to be deduced from a single on-chain signature; 374 addresses were stolen. The defect originated from an unaudited third-party component introduced two weeks before the incident.

• THORChain (May 15, approximately 10.7 million USD): Vault private keys were co-managed by multiple nodes (threshold signatures), intending that no single node holds the complete private key; but a malicious node gradually collected key fragments during multiple rounds of normal signing processes, ultimately reconstructing the complete private key.

From cross-chain validation to cloud keys and signature algorithms, these cases present different forms of the same failure: as long as there is a node on the critical path that can independently decide the destination of funds, it will become the attacker's preferred target.

Based on the same judgment, OKX not only verifies "whether the transaction was initiated by the user" in signature risk control but focuses more on letting users understand the transaction consequences before signing.围绕 durable nonce, account ownership changes, and other high-risk scenarios, OKX intercepted or alerted on related high-risk operations over 4 million times cumulatively in the first half, protecting approximately 526 million USD cumulatively, and has parsed over 50,000 on-chain methods, translating hard-to-read calldata into "which assets this transaction will change and what permissions will be granted." It needs to be pointed out that such capabilities can only reduce rather than eliminate risk; trading platforms and wallet service providers themselves are the highest value attack targets, as evidenced by the 2025 Bybit incident which was caused by breaching signature tools; no party can claim immunity.

User Side: Attacks Start at the Most Familiar Entry Points

When the cost of attacking protocols rises, attackers turn to users. According to the OKX report, attack entry points are now mostly scenarios where users are least guarded: app stores, top search results, friend accounts, conference software, and recruitment processes.

• Malicious browser extensions adopt a "local shell holding, cloud poisoning" model: the extension body contains no malicious logic and can pass static review by app stores; the real phishing pages are issued in real-time by remote servers and can change domains at any time (SlowMist captured a similar extension impersonating TronLink in May). Once users enter mnemonic phrases on such pages, asset control is handed over.

• Search ad poisoning is closer to daily operations. In a case recorded by the OKX report, a user searched for development tools after purchasing a new computer, clicked on the paid ad at the top of the search results, and executed an "installation command" in the terminal as prompted by the page; this command actually deployed a clipboard hijacking trojan; subsequently, when the user transferred approximately 20,000 USD, the recipient address was automatically replaced. The difficulty in defending against such attacks lies in: the victim only performed routine operations such as "searching for the official website, downloading tools, copying commands" throughout the process.

• Recruitment interview scams follow a "reconnaissance — profiling — targeted strike" path: attackers use technical interviews as a pretext, requiring applicants to share screens and open wallets to "confirm DeFi usage experience," actually recording their wallet addresses, holdings, and commonly used protocols; subsequently, they forge airdrop pages and deploy customized phishing information targeting protocols the victim has actually used. In a case recorded by the OKX report, the victim lost approximately 88,000 USD due to this.

There are two other new methods worth noting. One is fake "2FA security verification": attackers send emails in the name of the wallet official, using impersonating domains differing from the official one by only one character, supplemented by countdowns to create urgency, inducing users to enter mnemonic phrases to "complete verification." The other is business process fraud: using business scenarios such as "external audit" or "token vesting confirmation" as bait to deliver malicious attachments (a sample analyzed by SlowMist disguised scripts as documents with double extensions); after opening, it disguises itself as a system update to steal passwords, then applies for camera, screen recording, and keyboard listening permissions; its ultimate target is often office terminals and cloud service permissions, rather than personal wallets.

For the former, both reports give a consistent bottom-line reminder:

Any page requiring input of mnemonic phrases for verification, authentication, recovery, or upgrade is a scam. Mnemonic phrases are not verification codes, but asset control itself; no legitimate wallet will solicit mnemonic phrases via web pages.

Since many losses do not start with on-chain transactions but with earlier malicious applications or phishing websites, OKX has moved protection forward to devices and access entry points: its security scanning assistant has completed over 200,000 risk detections cumulatively, discovering over 60,000 high-risk applications; targeting phishing websites and malicious DApps, it has intercepted over 7 million risky website accesses cumulatively, striving to complete interception before users enter mnemonic phrases or connect wallets.

And before attackers contact users, a significant portion of risks are already buried in upstream links — the software supply chain.

Supply Chain: What Is Poisoned Is Not Software Packages, But Trust Itself

The specificity of supply chain poisoning lies in: victims make no mistakes, simply installing a dependency, updating a version, or visiting an official domain as usual. This is the chapter SlowMist's report focuses on most heavily; three representative cases reflect the evolution of such attacks in breadth, depth, and height.

• Breadth: Shai-Hulud Worm. In mid-May, an account released 637 malicious versions involving 317 package names within 22 minutes, affecting popular components such as echarts-for-react (monthly downloads over 3.8 million) and size-sensor (monthly downloads over 4.2 million). Malicious packages trigger obfuscated payloads during installation, systematically collecting AWS, GCP, Azure cloud credentials, Kubernetes cluster keys, SSH private keys, and other sensitive information for encrypted exfiltration; simultaneously, built-in self-propagation modules achieve worm-like diffusion, and persistence mechanisms targeting Claude Code and VS Code are pre-installed. Its theft target is no longer a single software package, but the permissions of the entire development pipeline.

• Depth: Layered transmission of the chain of trust. In March, the Python library LiteLLM, with monthly downloads reaching 97 million, was attacked. Attackers did not directly invade the library but first polluted Trivy, the security scanning tool relied upon by its build process, using this to steal release keys and push malicious versions. Developers trust LiteLLM, LiteLLM trusts Trivy, but Trivy has been breached, turning a security tool that should bear protection responsibilities into a link in the attack chain. Apifox's official CDN poisoning monitored by SlowMist in the same month belongs to the same category: attackers tampered with official scripts, implanting malicious code with random timers, which 发作 after a delay of half an hour to three hours after installation, making it difficult for users to associate anomalies with specific operations. This indicates that "official source" is no longer fully equivalent to "secure."

• Height: Attackers' "attack surface thinking." SlowMist Chief Information Security Officer 23pds warned in April: Lazarus sub-organization HexagonalRodent used high-salary remote positions as bait to induce developers to run code containing backdoors. According to the OKX report, the organization heavily leveraged ChatGPT and Cursor to generate code and scripts, used AI website building tools to forge company official websites and executive identities, and even utilized AI to "self-check" its own malicious code to evade detection; it stole wallet data from over 2,700 developer systems in just one quarter. Meanwhile, AI-generated code itself is creating new risk surfaces; according to OKX report citing OtterSec data: Georgia Tech attributed 35 of 74 CVEs in March to AI-generated code; a scan targeting approximately 1,400 "casually generated" applications discovered 2,038 severe vulnerabilities and over 400 exposed keys. "Functional availability" is not equivalent to "ready for production environment."

Similar incidents include node-ipc, which suddenly appeared with abnormal versions after 21 months of discontinuation, and TrapDoor, which coordinated poisoning across three major ecosystems. Their common pointing leads to a conclusion: in the past, the focus of protection was auditing code; now, the source of trust must be audited. This point is especially critical for the Agent ecosystem, as Agents will actively read, install, and execute external plugins. Based on this, OKX established admission audits and regular inspection mechanisms covering code security, permission scope, and external dependencies for plugin access in Agentic Wallet, moving the judgment of "whether dependencies are trustworthy" forward to before the plugin goes live.

AI: From Forging Single Content to Synthesizing Entire Environments

Connecting the previous content, we can see an implicit main line: there is natural trust in "smarter systems" from the outside. And the next collective misjudgment forming is: Agents are more reliable than humans. This cognition has not yet been corrected by enough accidents, but the cost has already manifested.

The May Bankr event is a case nearly unlike an "intrusion": attackers did not steal private keys, attack contracts, or invade servers; they just asked Grok to translate a piece of Morse code. BankrBot is an AI Agent deployed on the X platform that can automatically execute on-chain transactions based on natural language instructions. According to SlowMist's review, attackers first airdropped a member NFT to the associated wallet, triggering its high-permission mode, then sent a Morse code message asking Grok to translate; after Grok completed decoding and @BankrBot, it output the hidden transfer instruction (approximately 3 billion DRB tokens) in plaintext, and BankrBot treated it as a trusted instruction and executed it directly. Losses were approximately 150,000 to 200,000 USD (SlowMist recorded approximately 175,000 USD), and approximately 80% was recovered afterwards. Neither Grok nor BankrBot had vulnerabilities in the traditional sense; what truly failed was the unverified trust boundary between two automated systems. When Agents possess signing and transfer capabilities, a malicious instruction disguised as normal input can convert into a real asset loss.

SlowMist divides attacks targeting Agents into four layers according to their chain from cognition to execution, each corresponding to real cases:

• Input Layer: Indirect prompt injection. Embedding malicious instructions into external content that the Agent will actively read; the Bankr event belongs to this category.

• Memory Layer: Memory poisoning. Verification by Palo Alto Networks Unit 42 shows that malicious rules can be written into the Agent's long-term memory (such as marking the attacker's address as a "common收款 address"), and continue to take effect in subsequent sessions.

• Tool Layer: Permission penetration. In the OpenClaw "Claw Chain" event disclosed in May, security company Cyera discovered four vulnerabilities that could be chained (highest risk rating CVSS 9.6), which could be gradually chained into complete control over the execution environment.

• Supply Chain Layer: OpenClaw plugin center lacks auditing; malicious skills are disguised as function description documents, hiding commands via encoding, and adopt a "two-stage" method to dynamically pull the real malicious payload from remote.

AI's deeper change to phishing scenarios is upgrading fraud from "forging single content" to operating a whole set of "synthesized reality." According to the OKX report recounting Check Point's disclosure of the "Truman Show" operation, attackers lured victims into a private investment group where "experts" and "investors" within the group were all played by AI, continuously publishing analysis, displaying returns, and interacting in real-time; what the victim faced was not a single scammer, but a forged environment with a complete social and trust structure. The Singapore video conference case mentioned at the beginning of this article (AI-generated Prime Minister, President, and Monetary Authority representatives present simultaneously, loss of approximately 4.9 million SGD) is an extreme form of the same technique.

Thus, the attack target has undergone a fundamental shift: what needs to be deceived is no longer the user's judgment on a single piece of information, but their cognition of the authenticity of the overall environment. The judgment that "vulnerabilities appear in people" should also be further extended; when machines begin to make judgments instead of people, the risk boundary expands from "people" to all links that are defaulted to trust, including trust boundaries between machines. Traditional "identify suspicious information" style defense is failing.

In response, the industry has formed a consensus: Agents can organize transactions more intelligently, but their execution links must be subject to stricter constraints. SlowMist embeds security capabilities into the Agent operation flow; its MistEye Security Gate provides a "detect first, execute later" hard blocking mechanism for coding agents like Claude Code and Cursor; MistTrack Skills relies on over 400 million on-chain address labels covering 19 mainstream public chains to complete anti-money laundering risk screening before transfers. OKX's thinking is similarly not to entrust security entirely to the Agent's own judgment: regardless of how a transaction is generated, an independent risk detection will be performed before signing and broadcasting; once judged to exist risk, secondary confirmation is required or direct interception occurs.

VI. Fund Flow: Only About 12% Recoverable, Money Laundering Highly Industrialized

Returning to the opening numbers. Anti-money laundering and regulatory data in this chapter mainly come from SlowMist's report tracking: in the first half, only 18 incidents achieved fund recovery or freezing, involving approximately 389 million USD stolen, with approximately 118 million USD recovered or frozen, accounting for 12.3% of total losses. The vast majority of stolen funds are difficult to recover once transferred out; the fundamental reason is that the downstream money laundering link has become highly industrialized.

Taking Lazarus as an example, its money laundering chain is interlocked: first cutting off on-chain associations through privacy protocols like Umbra, then exchanging for Bitcoin via non-KYC cross-chain protocols like THORChain (liquidity is abundant and easy to split), subsequently dispersing funds to thousands of new addresses, processing through mixing services like Sinbad, YoMix, and Wasabi, and finally exchanging offshore for fiat via Chinese OTC counters or Russian exchange networks (such as Garantex and its successor Grinex). If stolen assets are tokens like rsETH, they will first be mortgaged in DeFi protocols to borrow real assets before exchange; KelpDAO stolen funds followed this path. Once funds enter cross-chain and mixing links, the recovery window basically closes.

Capital inflows into privacy protocols constitute side verification: according to SlowMist's statistics based on Dune, major privacy protocols cumulatively received inflows of approximately 974 million USD in the first half, of which Tornado Cash accounted for about 71%, Railgun accounted for about 23%, and nearly 90% of its inflows were stablecoins. The other end of the black industry is reflected in the "servitization" of tools: in the Drain-as-a-Service platforms tracked by SlowMist, Rublevka Team has cumulatively stolen over 10 million USD since 2023, completing at least 240,000 successful thefts, and providing up to 80% splits to downstream; StepDrainer prices the whole set of phishing tools clearly, with complete source code at approximately 750 USD. The acquisition threshold for attack tools has dropped to an extremely low level.

Regulatory constraints are also tightening and beginning to precisely point at key nodes of money laundering channels. In the dynamics sorted by SlowMist, there are three items most directly acting on the above links: the EU's 20th round of sanctions against Russia comprehensively prohibited transactions with crypto service providers within Russia for the first time, directly pointing at cash-out channels like Garantex and Grinex; Dubai DIFC comprehensively prohibited privacy coin transactions, targeting the mixing link; China's eight departments' new regulations in February explicitly stated for the first time that issuing RMB-pegged stablecoins offshore without approval is not allowed. Other measures also advance in the same direction; Hong Kong issued the first batch of stablecoin issuer licenses; South Korea plans to fine exchanges hacked up to 10% of the stolen amount; the US GENIUS Act landed the first stablecoin rules. The starting point of the money laundering link is often a malicious address or a high-risk authorization; the OKX report also emphasizes that its KYS risk capability will continue to identify abnormal addresses, malicious tokens, and high-risk authorizations, striving to complete risk marking before funds enter mixing and cross-chain channels.

VII. Conclusion: Move the Defense Line Forward Before Confirmation

At the moment the Drift signer pressed the confirm button, all defenses had lost meaning. Most lessons in the first half of 2026 occurred before the action of "confirmation": someone signed a transaction they did not understand, some system granted excessive permissions, some team trusted the wrong data source, and some user believed a disguise 逼真 enough.

Whether it is OKX's proposal of "understand before signing, avoid single-point risks, verify before trusting," or SlowMist's "verify first, trust later" thinking throughout, both ultimately point in the same direction: move judgment forward before the action occurs. Based on this, it can be summarized into three lists.

For Projects: Critical transactions should not only confirm "who signed," but also confirm "whether the signer understands the consequences." Multisig, contract upgrades, Owner changes, cross-chain configurations, and other operations should have parseable, simulatable, and reviewable processes; large fund outflows and permission changes should set timelocks and abnormal alerts. No single link should be allowed to independently decide the destination of funds, such as an RPC, a validation node, a cloud key, or an admin address. Dependency scanning should be incorporated into normalized processes; requests involving funds, permissions, and deployment should not be confirmed via a single channel only; video conferences and friend accounts do not constitute identity proof.

For Ordinary Users: Transactions that cannot be understood should not be signed. Large assets and daily interaction wallets should be stored separately; unused authorizations should be revoked regularly; infinite allowance authorizations, Permit/Permit2, and other operations require particular caution; most attacks do not transfer assets immediately but first obtain authorization, then wait for an opportunity. Any page soliciting mnemonic phrases is a scam; any request requiring remote control, screen sharing, installing unfamiliar conference software, or executing terminal commands should be verified first.

For Agent Users and Developers: Input filtering, tool call sandboxes, permission grading, secondary confirmation for sensitive operations, and pre-execution verification are basic configurations for any Agent capable of holding funds, not options. System prompts should not be treated as security control means; the Agent should be responsible for generating schemes, while an inflexible rule system is responsible for authorization.

The first half of 2026 was not an overall improvement in security 形势, but a systemic migration of attack paradigms, from contract code to people, infrastructure, and trust itself outside the code. In an era where synthesized reality becomes possible, what needs verification is no longer a single piece of information, but the environment itself. The goal of security is never to require every participant to become an expert, but to intercept risks before confirmation at the links where they are most prone to error.

What needs repairing most in the first half of 2026 is not a certain contract, but people and processes themselves.