Anthropic Says AI Can Hack Smart Contracts After Spotting $4.6M in Exploits
Anthropic has proven that highly effective AI methods can discover weaknesses in blockchain apps and switch them into worthwhile assaults value tens of millions of {dollars}, elevating contemporary issues about how uncovered DeFi actually is.
In a recent study with MATS and Anthropic Fellows, the corporate examined AI brokers on a benchmark known as SCONE-bench (Smart CONtracts Exploitation), constructed from 405 sensible contracts that have been really hacked between 2020 and 2025.
When they ran 10 main fashions in a simulated atmosphere, the brokers managed to use simply over half of the contracts, with the simulated worth of stolen funds reaching about $550.1m.
To cut back the prospect that fashions have been merely recalling previous incidents, the workforce then appeared solely at 34 contracts that have been exploited after March 1, 2025, the newest data cutoff for these methods.
Opus 4.5 And GPT-5 Located $4.6M In Value From New Exploit Targets
On that cleaner set, Claude Opus 4.5, Claude Sonnet 4.5 and GPT-5 nonetheless produced working exploits on 19 contracts, value a mixed $4.6m in simulated worth. Opus 4.5 alone accounted for about $4.5m.
Anthropic then examined whether or not these brokers might uncover model new issues reasonably than replay outdated ones. On Oct. 3, 2025, Sonnet 4.5 and GPT-5 have been run, once more in simulation, towards 2,849 just lately deployed Binance Smart Chain contracts that had no identified vulnerabilities.
Both brokers discovered two zero-day bugs and generated assaults value $3,694, with GPT-5 doing so at an API value of about $3,476.
Tests Ran Only On Simulated Blockchains With No Real Funds At Risk
All of the testing passed off on forked blockchains and native simulators, not stay networks, and no actual funds have been touched. Anthropic says the purpose was to measure what’s technically attainable at this time, to not intervene with manufacturing methods.
Smart contracts are a pure check case as a result of they maintain actual worth and run absolutely on chain.
When the code goes mistaken, attackers can typically pull property out straight, and researchers can replay the identical steps and convert the stolen tokens into greenback phrases utilizing historic costs. That makes it simpler to place a concrete quantity on the harm an AI agent might trigger.
SCONE-bench measures success in {dollars} reasonably than simply “sure or no” outcomes. Agents are given code, context and instruments in a sandbox and requested to discover a bug, write an exploit and run it. A run solely counts if the agent finally ends up with at the very least 0.1 further ETH or BNB in its steadiness, so minor glitches don’t present up as significant wins.
Study Shows Attack Economics Improve As Token Costs Decline
Over the previous yr, the examine discovered that potential exploit income on the 2025 issues roughly doubled each 1.3 months, whereas the token value of producing a working exploit fell sharply throughout mannequin generations.
In follow, which means attackers get extra working assaults for a similar compute funds as fashions enhance.
Although the work focuses on DeFi, Anthropic argues that the identical expertise carry over to conventional software program, from public APIs to obscure inside companies.
The firm’s core message to crypto builders is that these instruments reduce each methods, and that AI methods able to exploiting sensible contracts may also be used to audit and repair them earlier than they go stay.
The put up Anthropic Says AI Can Hack Smart Contracts After Spotting $4.6M in Exploits appeared first on Cryptonews.
