OpenAI's AI Red Team: The $108M Crypto Hack Benchmark

The core finding from EVMbench is stark: AI agents are far more effective at exploiting vulnerabilities than they are at finding or fixing them. This performance gap defines the new offensive threat landscape. The benchmark tested agents against a curated set of 120 potential weaknesses drawn from 40 smart contract audits, simulating real-world attack scenarios.

In detection, Anthropic's Claude Opus 4.6 scored the highest mean result. Yet even this top performer's success in identifying flaws remains limited. The data shows agents often stop after finding a single issue, failing to conduct a thorough audit. Their patching ability is similarly constrained, struggling to maintain full functionality while removing subtle vulnerabilities.

The most significant performance was in exploitation. Here, agents excel at the explicit, iterative task of draining funds. OpenAI's GPT-5.3-Codex more than doubled the effectiveness of an earlier model in this setting. This creates a clear asymmetry: the tools to attack are advancing faster than the tools to defend.

The implication is a lowered barrier to entry for attackers. As AI-driven hacks continue to flow, with over $108 million in hacks and exploits recorded in 2026, the market for robust defensive tools becomes critical. The benchmark quantifies a dangerous edge for offense, making automated security solutions a necessity, not a luxury.

The Attack Losses: Real-World Cost of the Gap

The security gap quantified by EVMbench translates directly into real financial losses. The total cost is already severe, with over $108 million in hacks and exploits recorded in 2026. This isn't abstract risk; it's a tangible drain on capital that accelerates the need for better defenses.

Recent high-profile attacks illustrate the scale. In February, the DeFi protocol Moonwell lost nearly $2.7 million in crypto due to a misconfigured oracle that undervalued a key asset by 99.9%. The platform was left with approximately $1.78 million in bad debt. Just days earlier, the cross-chain bridge CrossCurve suffered a $3 million exploit through a flaw in its ReceiverAxelar contract, which attackers used to bypass validation checks.

These incidents align with a broader, systemic trend. A 2025 security report found that 83% of Web3 losses stemmed from access control issues and infrastructure failures, not just code flaws. This means the vulnerabilities AI agents are best at exploiting-like flawed validation logic or misconfigured systems-are the same ones causing the majority of financial damage. The cost is mounting, and the attack surface is expanding beyond pure code.

The Market Flow Implications: Defensive AI as a New Asset Class

The primary investment risk is clear: offensive AI tools will be deployed by attackers faster than defensive solutions can be scaled. The EVMbench data shows agents are best at exploiting vulnerabilities, with a clear asymmetry favoring attack. This creates a persistent, widening gap that will keep exploit costs elevated until defensive AI catches up.

Key watchpoints are the adoption rate of defensive AI auditing tools and the trajectory of new exploit losses. The market is signaling urgency. OpenAI and Paradigm's joint launch of EVMbench, which draws on 120 curated vulnerabilities from 40 audits, is a direct response to recent attacks like Moonwell. The tool's release came just days after a bug in AI-generated code cost Moonwell users nearly $2.7 million. This timing frames defensive AI not as a future possibility but as an immediate, capital-intensive need.

Viewed through a flow lens, defensive AI is emerging as a new asset class. The $108 million in 2026 hacks represents a massive, recurring cost that must be mitigated. As protocols and exchanges seek to reduce this drain, they will allocate capital to AI-powered security solutions. The market flow here is twofold: capital will exit vulnerable protocols and flow into the defensive tools that can audit and patch code faster than AI attackers can find new exploits. This is a high-stakes race where the first-mover advantage in defensive AI could capture significant market share.