Submission + - AI Agents Casually Rob $4.6 Million (In Simulation), Cost $3,476 To Run (anthropic.com)
weiyeh writes: Remember when we worried about script kiddies? Meet their AI-powered, profit-optimizing successors. Researchers from Anthropic and the MATS program have released SCONE-bench, a benchmark showing that AI agents can now autonomously exploit blockchain smart contracts—and they're getting scary good at it. From the research report:
"On contracts exploited after March 2025 (beyond the models' training data), Claude Opus 4.5, Claude Sonnet 4.5, and GPT-5 developed working exploits worth a collective $4.6 million in simulated stolen funds. The top performer, Opus 4.5, successfully compromised 50% of these contracts."
"On contracts exploited after March 2025 (beyond the models' training data), Claude Opus 4.5, Claude Sonnet 4.5, and GPT-5 developed working exploits worth a collective $4.6 million in simulated stolen funds. The top performer, Opus 4.5, successfully compromised 50% of these contracts."
