Researcher shows major AI agent benchmarks can be easily gamed

A Berkeley researcher demonstrated that popular AI agent benchmarks—used to measure progress in autonomous systems—have fundamental flaws that let models score high without genuine capability. This matters for indie makers building on these benchmarks: the leaderboard scores you're chasing may not reflect real-world performance, and the supposedly "state-of-the-art" agents you're comparing against might just be exploiting measurement loopholes.