AI agents becoming professional hackers? Anthropic reveals that GPT-5 and Claude can replicate on-chain attack patterns.

Anthropic's latest research indicates that AI's capability for cyber attacks is rapidly approaching that of human hackers. Several models, including GPT-5 and Claude Opus, have been able to replicate over 200 historical smart contract attacks and even discover previously undisclosed zero-day vulnerabilities on the BNB Chain. Experts warn that the low cost, high efficiency, and automation characteristics of AI will allow attack behaviors to be replicated on a large scale, which also means that the rhythm of offensive and defensive cybersecurity in Web3 will face a comprehensive reshaping.

The Era of AI Hackers Has Arrived: 10 Major Models Successfully Replicate Over 200 On-Chain Attacks

Anthropic announced security assessments for 10 cutting-edge AI models, including GPT-5, Claude Opus 4.5, DeepSeek V3, and Llama 3.

Research shows that AI has successfully recreated 207 out of 405 historical attacks, simulating a loss of 550 million dollars, and performing comparably to skilled human hackers in more than half of the cases. This indicates that the entire process of attacks and the logic of exploiting vulnerabilities have been fully understood and executed by AI.

From Reproduction to Discovery: GPT-5 and Claude Can Uncover Zero-Day Vulnerabilities in Protocols

In the more challenging “zero-day vulnerability (zero-day vulnerability)” test, Anthropic had the model scan 2,849 new contracts extracted from 9.4 million contracts on the BNB Chain.

The results were surprising, as Claude Sonnet 4.5 and GPT-5 each found two undisclosed vulnerabilities and successfully generated approximately $3,694 in simulated profits in a sandbox environment. This indicates that AI is not only capable of mimicking attacks but also possesses the ability to “actively discover” unknown weaknesses.

( Chinese hackers use AI agents to autonomously invade 30 global institutions, Anthropic issues urgent warning )

Why are AI attacks easy to spread? Open source culture is the key.

Regarding the reason why the scale of AI attacks can easily amplify quickly, SovereignAI COO David Schwed explained that due to the abundance of publicly available audit and vulnerability reports in the market, along with the presence of forked versions among various projects, the advantage of AI rapidly reviewing and comparing them can further magnify the situation:

A simpler approach is to find a disclosed vulnerability and then identify similar projects based on its architecture, and attempt to exploit the vulnerability, as it may not have been patched yet (patch).

Once the model's capabilities continue to improve, attacks will become replicable, and at that time will also move towards automation and scaling, with the speed gap between attackers and defenders only growing larger.

Defenders can also arm themselves with AI: attacks are faster, and repairs must be faster too.

Anthropic emphasizes that the capabilities of AI in using tools, error correction, and long-term task planning are rapidly evolving, significantly increasing attack speed and success rates; however, the same advancements can also be applied to the defense side.

Schwed believes that as long as rigorous testing, real-time monitoring, and interval blocking mechanisms are adopted, AI can actually put white hats and attackers on the same starting line. He emphasized: “What bad people can do, good people can do as well. We just need to think in a different way.”

( The FBI warns that AI-driven “cyber-physical attacks” are on the rise, posing a significant threat to infrastructure )

This article discusses whether AI agents can become professional hackers. Anthropic reveals that GPT-5 and Claude can replicate on-chain attack patterns, which first appeared in Chain News ABMedia.