AI Agents Demonstrate Alarming Ability to Exploit Smart Contracts, Potentially Draining Millions

In a sobering new study from AI safety leader Anthropic , autonomous AI agents have proven capable of identifying and exploiting vulnerabilities in blockchain smart contracts at a scale that could enable multimillion-dollar thefts.

Researchers developed a novel benchmark called SCONE-bench, comprising 405 real-world smart contracts that were successfully exploited between 2020 and 2025 across major chains like Ethereum, BNB Smart Chain, and Base. Ten frontier AI models - including Anthropic's Claude Opus 4.5 and Claude Sonnet 4.5 , OpenAI's GPT-5, and others such as DeepSeek V3 - were tasked with autonomously analyzing the code, crafting exploit scripts, and executing attacks in a simulated environment using tools like Python, Foundry, and forked blockchains.

Even here, leading models like Opus 4.5, Sonnet 4.5, and GPT-5 succeeded on 19 cases (55.8%), with simulated damages reaching $4.6 million. Opus 4.5 alone accounted for $4.5 million, demonstrating superior reasoning in high-value scenarios.

Performance varied notably between models, underscoring their differing approaches to problem-solving. In one instance, GPT-5 extracted $1.12 million from a vulnerable contract, while Opus 4.5 maximized the same flaw to drain $3.5 million by optimizing transaction sequencing and liquidity routing.

Both independently discovered two previously unknown zero-day vulnerabilities, yielding $3,694 in simulated profits at a total inference cost of just $3,476 for GPT-5 (roughly $1.22 per contract scanned).

Anthropic observed explosive progress in AI exploitation capabilities: over the past year, simulated exploit revenue on post-March 2025 contracts roughly doubled every 1.3 months, while compute costs plummeted - dropping 70% in six months for some generations. This Moore's Law-like acceleration means more than half of 2025's real-world DeFi hacks, traditionally executed by elite human teams, could now be automated by off-the-shelf AI agents.

The implications extend far beyond crypto. Smart contracts' public, immutable nature makes them an ideal proving ground, but the underlying skills - reading complex code, tracing execution paths, and chaining transactions - apply equally to traditional software vulnerabilities, APIs, and infrastructure.

Yet the report strikes a balanced tone: the same autonomous agents that uncover exploits can serve as powerful defensive tools. Anthropic plans to open-source the SCONE-bench dataset, enabling developers to stress-test contracts pre-deployment and integrate AI-driven auditing into standard practices.

As one researcher noted, the race is on - those who deploy AI for proactive security first will hold a decisive edge in an era where attackers never sleep, tire, or forget. For the DeFi industry, already reeling from billions in cumulative losses, this study serves as a urgent wake-up call to embrace AI not just as a threat, but as the frontier in blockchain resilience.