OKX has published its H1 2026 Web3 Security Semi-Annual Report, revealing a sharp divergence between the frequency and financial impact of crypto security incidents. The exchange recorded 182 publicly disclosed security events during the first half of 2026, a 50% increase from 121 incidents in H1 2025. However, total losses fell approximately 60% to $956 million, down from $2.373 billion in the same period a year earlier. The report adds to a growing body of industry data from firms such as CertiK, Immunefi, and Chainalysis that have tracked Web3 security trends, providing a benchmark at a time when protocols and platforms face an evolving threat landscape.

The report was jointly produced by OKX, SlowMist, and OtterSec.

182 Incidents and $956M in Losses

The steep decline in losses was primarily attributed to the absence of a single outlier event comparable to the approximately $1.5 billion supply-chain attack on Bybit in February 2025. The higher incident count alongside lower total losses means the average loss per event decreased compared to H1 2025, when the Bybit event alone accounted for a significant share of total losses.

According to OtterSec's compilation of the ten largest project-level losses, KelpDAO led with approximately $292 million lost, followed by Drift Protocol at approximately $285 million.

Attack Patterns Shift Away from Smart Contract Exploits

The report highlights a notable shift in attack vectors. The most significant financial losses no longer originate primarily from smart contract vulnerabilities. Instead, losses were driven by supply chain compromises, social engineering, cloud key theft, and single-point validation failures. This trend aligns with observations across the industry that as smart contract auditing and formal verification practices have matured, attackers have increasingly targeted operational infrastructure, human trust, and key management rather than protocol logic itself.

Supply chain attacks accounted for approximately $298 million in losses, followed by contract logic issues at $152 million and private key leaks at $130 million.

AI Agents Manipulated in Novel Attacks

The report cited the Bankr incident as an example of how AI agents can be exploited. Attackers activated an Agent membership privilege and transmitted a Morse code prompt injection to xAI's Grok, which decoded the message and forwarded it to @bankrbot. The bot accepted the instruction as trusted input and executed transfers valued at approximately $150,000 to $200,000 on the Base network. As AI agents gain adoption across DeFi, wallet interfaces, and automated trading workflows, such attack vectors represent an emerging frontier that extends beyond traditional smart contract risks.

AI has also broadened the crypto attack surface through social engineering, identity forgery, and automated code generation. The report referenced North Korea's HexagonalRodent operation, which leveraged ChatGPT, Cursor, AI-generated websites, and fraudulent recruitment campaigns to compromise over 2,700 developer systems. OtterSec additionally linked 35 of 74 CVEs recorded in March 2026 to AI-generated code, a finding with implications for development practices as AI coding assistants see widespread adoption across the software industry.

Users and AI Agents Emerge as Primary Targets

As users and AI agents become the principal targets of attack, the report anticipates that the next phase of Web3 security will center on proactive, layered defenses that intervene earlier in the attack chain — before signing, execution, or trust is granted.

The report expects wider adoption of sandboxed tool calling, permission tiering, pre-execution simulation, and secondary human confirmations for sensitive transactions. It also anticipates that more wallets and platforms will deploy real-time device scanning, URL and DApp risk detection, and malware identification to intercept threats before seed phrases are exposed or wallet connections are established.

The report advises users to understand transactions before signing, eliminate single points of failure, and verify before trusting. Recommended measures include rigorous audits of AI-generated code, scoped permissions, transaction transparency, and layered safeguards that do not depend solely on AI judgment.

Source: OKX Web3 Security Report H1 2026