AI models not only analyze code — they can already hack it. The test results from Anthropic are alarming for the entire DeFi.
AI generated exploits for more than half of the tested contracts.
The models discovered two previously unknown vulnerabilities — Zero-Day.
SCONE-bench: the first AI test that measures losses in dollars
Anthropic created the SCONE-bench benchmark to assess the ability of AI models to identify and exploit vulnerabilities in smart contracts. The results are concerning.
In a test of 405 contracts hacked between 2020 and 2025, the models successfully executed attacks on 207 of them, generating $550.1 million in simulated losses.
In the second test — involving only contracts that were hacked after the models' knowledge cutoff date — AI still created functioning exploits, generating $4.6 million in losses.
The most alarming was the Zero-Day test: the GPT-5 and Sonnet 4.5 models discovered entirely new vulnerabilities, generating exploits worth $3694 — with an API cost of $3476, meaning that attacks are already profitable.
According to Anthropic, AI capabilities double every 1.3 months, while the cost of attacks decreases. This means that DeFi is entering a new era of threats.