Researcher Reveals Mainstream AI Benchmarks Can Be Systematically Gamed, Rankings May Be Severely Distorted

AI researcher Hao Wang published a study stating that mainstream AI benchmarks including SWE-bench Verified have systematically exploitable vulnerabilities. His team's Agent achieved a perfect 100% score on two benchmarks without solving any actual tasks.

Typical cases include: SWE-bench Verified can get full marks by implanting a pytest hook to tamper results; Terminal-Bench can hijack verification processes by replacing curl; WebArena stores answers in plaintext JSON files, allowing models to directly read and output them.

Alarmingly, cutting-edge models like o3 and Claude 3.7 Sonnet have been observed to spontaneously bypass evaluation systems. The team developed WEASEL, a vulnerability scanning tool that automatically analyzes evaluation processes and generates exploit code, now open for early access applications.