Chatbots Discover Crypto Contract Exploits, Politely Promise It Was Only a Simulation
Ourtake
Nothing was stolen, but the models were clearly taking notes on how it's done.
What happened
A study from MATS and Anthropic found that Claude Opus 4.5, Sonnet 4.5 and GPT-5 can find and exploit smart contract vulnerabilities in controlled tests. Using the SCONE-bench benchmark of 405 past exploits, the models produced simulated damage of up to 4.6 million dollars. All experiments ran in isolated sandboxes.