Frontier AI Models Demonstrate Human-Level Capability in Smart Contract Exploits
AI Models Reach Human-Level Capability in Smart Contract Exploits
In a groundbreaking study, Anthropic has demonstrated that AI models can now match the performance of skilled human attackers in more than half of smart contract exploits recorded on major blockchains over the last five years. This research is particularly significant as it highlights the rapid advancement of artificial intelligence in cybersecurity threats.
- Anthropic tested ten AI models on 405 historical smart contract exploits and reproduced 207 of them.
- Three models generated $4.6 million in simulated exploits on contracts created after their training cutoff.
- Agents also discovered two new zero-day vulnerabilities in recent Binance Smart Chain contracts.
The findings, published Monday by Anthropic, underscore the potential of AI in automating and weaponizing cybersecurity threats. Notably, the company tested ten models including Llama 3, Sonnet 3.7, Opus 4, GPT-5, and DeepSeek V3 on a dataset of 405 historical smart contract exploits. The agents produced working attacks against 207 of these contracts, totaling $550 million in simulated stolen funds.
Experts have noted that the ease with which AI can exploit vulnerabilities highlights the growing accessibility of these flaws. David Schwed from SovereignAI highlighted that many of these vulnerabilities are already publicly disclosed and learnable by AI systems, making them easily exploitable against existing smart contracts.
The research also reveals how quickly automated systems can identify new zero-day vulnerabilities. For instance, Anthropic’s Claude Sonnet 4.5 and GPT-5 models uncovered two undisclosed flaws that produced $3,694 in simulated value each. The company emphasized that all tests ran in sandboxed environments to replicate blockchains without impacting real networks.
The study further highlights improvements across different model generations, attributed to advances in tool use, error recovery, and long-horizon task execution. Across four generations of Claude models, token costs fell by 70.2%, indicating the potential for increased efficiency as AI technologies advance.
Anthropic’s research demonstrates that the capabilities enabling agents to exploit smart contracts also apply to other types of software. As costs fall, the window between deployment and exploitation is shrinking, urging developers to adopt automated tools in their security workflows to match these advancements.
In conclusion, while Anthropic's findings present a concerning outlook on AI-driven cyber threats, Schwed suggests that proper controls and rigorous testing can mitigate many of these risks. Developers must remain vigilant as the threat landscape evolves with technological advancements.
Comments
Post a Comment