The ability of AI agents to exploit smart contracts in Ethereum and other Blockchains has raised urgent concerns about the potential economic risk of such autonomous cyber capability.
You can read more about it here:
- Frontier AI models such as GPT-5 and Claude exploited Ethereum and other Blockchains for smart contracts in simulated test.
- The AI models discovered previously unknown security flaws—called zero-day vulnerabilities—in software (in this case, smart contracts on Ethereum).
- The findings highlight the need for AI-powered defensive strategies that are proactive, since AI agents can now identify blockchain exploits as profitable as human hackers.
A joint project By Anthropic, MATS Fellows tested AI models on 405 real world contracts that were exploited in the period between 2020 and 25 using the Smart CONtracts Exploitation Benchmark (SCONE-bench).
Claude Opus 4.5, Claude Sonnet 4.5, and GPT-5 all produced exploits totaling $4.6 million in simulated attack simulations on contracts exploited by AI after March 2025. These results demonstrate a tangible lower bound for the financial harm AI may cause. GPT-5, and Sonnet, extended the test to include 2,849 contracts that had been deployed recently with no known vulnerability. They discovered two zero-day vulnerabilities which generated simulated profit of almost $3,700.
Scan-bench: Quantifying exploits using dollars instead of bugs
In contrast to traditional benchmarks, SCONE Bench evaluates AI exploits based on financial metrics, allowing for a better understanding of risk. Smart contracts work well for this strategy because they can translate vulnerabilities into money stolen, and the simulations enable researchers to calculate potential losses.
Ten AI models exploited 207 of the 405 SCONE-bench contracts and generated $550.1 in simulated stolen funds. Frontier models were able to successfully exploit contracts even after the knowledge cutoff date.
AI Exploitations: Examples Concrete
A token calculator on an Ethereum compatible contract was accidentally left writeable. AI repeatedly called function to increase token balance. This generated simulated profit of $2,500 Under peak liquidity conditions there is a potential for a potential $19,000. Later, an independent white-hat operation recovered the assets.
The research underscores that AI agents are now approaching human-level capability in tasks like control-flow reasoning, boundary analysis, and exploiting software vulnerabilities—a skill set directly applicable to blockchain and traditional software systems alike.
Study highlights the rapid acceleration of AI-based cyber capabilities, ranging from automated exploitations of blockchains applications to network intrusions. This tool allows developers of smart contracts to test their systems and stress them before they are deployed.
The researchers claim that the results are proof of concept for profitable real-world auto-exploitation. This highlights the need for AI-powered proactive defenses in order to protect digital assets and financial systems.
“This article is not financial advice.”
“Always do your own research before making any type of investment.”
“ItsDailyCrypto is not responsible for any activities you perform outside ItsDailyCrypto.”
Source: crypto.news
